%
AEOSR-TR- 9 1 0 ^ Q
ih-iS-,~.
REPORT DOCUMENTATION PAGE
*0pfW^
gM9 No. 070441$$
2. Riraiir oATt
5 June 1991
4. TITU ANO SUtmU
United States Air Force Sununer Faculty Research Projgram
i99o xjoVomp
Program Technical Report ’fw/ateh- 1 & ,3) _
r^rTniTr-TTfii
.S. rUNOWM NUDMIU
F49820>88-C-0053
Mr Rodney Darrah
4]mTT-r.T— T-.Trrrr^iT-rrmrrni
AND AOOfll
Universal Energy Systems (UES)
4401 Dayton -Xenia
Dayton OH 45432
I r ¥-rr»T^ n it-t
ftlKWruUMMi
AFOSR'TR. o T (i 0
i iMNi4«tN4/MdNltO«N4 AaiNCY NAMKS) AMO A
AFOSR/NI
Bldg 410
Bolling AFB DC 20332-6448
Lt Col V. Claude Cavender
13. AMTMACT (mmmimiCOwvMO
The United States Air Force Summer Faculty Research Program
(USAF-SFRP) is designed to introduce university, college, and
technical institute faculty members to Air Force research.
This is accomplished by the faculty members being selected on
a nationally advertised competitive basis for a ten-week
assignment during the summer intersession period to perform
research at Air Force laboratories/centers. Each assignment
is in a subject area and at an Air Force facility mutually
agreed upon by the faculty members and the Air Force. In
addition to compensation, travel and cost of living
allowances are also paid. The USAF-SFRP is sponsored by the
Air Force Office of Scientific Research,
14. SUlilCT niuM
91 1223 1^5
17. SfCUWTY OASStfOnON
or MMRT
11. SICUWTY OASSViCAnON
or THIS rAfii
UNCLASSIFIED
UNCLASSIFIED
ao. UMTATIONdf AiSTRAa
ir.i
UNITED STATES AIR FORCE
SUMMER FACULTY RESEARCH PROGRAM
1990
PROGRAM TECHNICAL REPORT
UNIVERSAL ENERGY SYSTEMS, INC.
VOLUME m of IV
Program Director, UES
Rodney C. Darrah
Program Administrator, UES
Susan K. Espy
Program Manager, AFOSR
Lt. Col. Claude Cavender
Submitted to
Air Force Office of Scientific Research
Bolling Air Force Base
Washington, DC
Ao*(w»aiou Por
HTIS" OSAikI
mic nii □
Ury?»Mao"ui«.ici Q
Just. Lifi cat -
ly -
Distribution/
Av-ftllabtiilv Codes
jAvaii tmd/or
Dist I Special
December 1990
TABLE OF CONTENTS
Section Pace
Preface . i
List of Participants . ii
Participant Laboratory Assignment . xxxv
Research Reports . xl
PREFACE
The United States Air Force Summer Faculty Resarch Program (USAF-SFRP) is designed
to introduce university, college, and technical institute faculty members to Air Force lesearch.
This is accomplished by the faculty members being selected on a nationally advertised
competitive basis for a ten-week assignment during the summer intersession period to perform
research at Air Force laboratories/centers. Each assignment is in a subject area and at an Air
Force facility mutually agreed upon by the faculty members and the Air Force. In addition to
compensation, travel and cost of living allowances are also paid. The USAF-SFRP is sponsored
by the Air Force OfHce of ScientiEc Research, Air Force Systems Command, United States Air
Force, and is conducted by Universal Energy Systems, Inc.
The specific objectives of the 1990 USAF-SFRP are:
(1) To provide a productive means for U.S. faculty members to participate in research
at Air Force Laboratories/Centers;
(2) To stimulate continuing professional associadon among the faculty and their
professional peers in the Air Force;
(3) To further the research objectives of the United States Air Force;
(4) To enhance the research productivity and capabilities of the faculty especially as
these relate to Air Force technical interests.
During the summer of 1990, 165-facuity members participated. These researchers were
assigned to 23 USAF laboratories/centers across the country. This four volume document is a
compilation of the final reports written by the assigned faculty members about their summer
research efforts.
List OF 1990 PARTICIPANTS
NAME / ADDRESS
DEGREE, SPECIALTY, LABORATORY ASSIGNED
Thomas. Abraham
Instructor
Saint Paul’s College
Dept of Science and Math
Lawrenceville, VA 23868
(804) 848-3111
Degree:
Soecialtv:
Assigned;
MS
Mathematics
Avionics Laboratory
Charles Alajajian
Assistant Professor
West Virginia University
PO Box 6101
Morgantown, WV 26506
(304) 293-6371
Specialty:
Assigned:
PhD
Electrical Engineering
Rome Air Development Center
Theodore Aufdemberge
Professor
Concordia College
4090 Geddes Road
Ann Arbor, MI 48105
(313) 985-7349
Degree:
Specialty:
Assigned:
PhD
Physical Chemistry
Geophysics Laboratory
Richard Backs
Assistant Professor
Wrig^t State University
Dept of Psychology
Dayton, OH 45435
(513) 873-2656
Degree:
Soecialtv;
Assigned;
PhD
Psychology
Aerospace Medical Research Lab.
William Bannister
Professor
Lowell, University of
Dept of Chemistry
Lowell, MA 01854
(508) 934-3682
Degree:
Soecialt'/;
Assigned;
PhD
Organic Chemistry
Engineering & Services Center
u
•h
y-
NAME / ADDRESS
DEGREE, SPECIALTY, LABORATORY ASSIGNED
Margaret Batschelet
Degree:
PhD
¥
Assistant Professor
Sbecialtv:
English
Texas-San Antonio, Univ. of
Assigned:
Human Resources Laboratory
Division of English
Training Systems
San Antonio, TX 78285
■
(513) 691-5357
1
Frank Battles
Degree:
PhD
Professor
Soecialtv:
Physics
1
Massachusetts Maritime Acad.
Assigned:
Geophysics Laboratory
Basic Science Dept.
>
Buzzards Bay, MA 02532
A
V
(508) 759-5761
m
John Bay
Degree:
PhD
>
Assistant Professor
SBSgiiltt;
Electrical Engineering
A
Virginia Polytech Institute
Dept of Electrical Eng.
Assigned:
Flight Dynamics Laboratory
i
Blacksburg, VA 24061
(703) 231-5114
*T
'.r
Reuben Benumof
j^QSs:
PhD
¥
Professor
SDe*.'ialtv:
Physics
\
Staten Island, College of
130 Stuyvesant PI.
Assigned:
Geophysics Laboratory
Staten Island, NY 10301
»■
(718) 390-7973
Phillip Bishop
Degree:
PhD
¥
Assistant Professor
Soecialtv:
Exercise Physiology
f
Alabama, University of
PO Box 870312
Assigned:
School of Aerospace Medicine
►
Tuscaloosa, AL 35487
%
(205) 348-8370
V
t'
>
¥
iii
NAME / ADDI^SS
DEGREE, SPECIALTY, LABORATORY ASSIGNED
Robert Blystone
Professor
Trinity University
715 Stadium Dr.
San Antonio, TX 78212
(512) 736-7243
Deeree:
Soecialtv:
Assigned:
PhD
Zoology
School of Aerospace Medicine
Michael Breen
Assistant Professor
Alfred University
Myers Hall
Alfred, NY 14802
(607) 871-2258
Degree;
Soecialtv:
Assigned:
PhD
Mathematics
Avionics Laboratory
Bruno Breitmeyer
Professor
Houston, University of
Dept of Psychology
Houston, TX 77204
(713) 749-6108
Degree:
Soecialtv:
Assigned:
PhD
Experimental Psychology
School of Aerospace Medicine
•
Mark Biusseau
Assistant Professor
Arizona, University of
429 Shantz Bldg. #38
Tucson, AZ 85721
(602) 621-3244
DS£SSI
Soecialtv:
Assigned:
PhD
Environmental Chemistry
Engineering & Services Center
David Buckalew
Assistant Professor
Xavier University
7325 Palmetto St.
New Orleans, LA 70125
(504) 483-7527
Degree:
Soecialtv:
Assigned:
PhD
Biology
Occupational & Environmental
Health Laboratory
IV
NAME / ADDRESS
DEGREE, SPECIALTY, LABORATORY ASSIGNED
in
Thecxiore Burkey
Decree:
PhD
J
Assistant Professor
Soecialtv:
Chemistry
Memphis State University
Chemistry Dept
Memphis, TN 38152
Assicned:
Frank J. Seiler Research Lab.
i.
k-
(901) 678-2634
i
Larry Byrd
PhD
Assistant Professor
Specialty:
Mechanical Engineering
Arkansas State University
PO Box 1740
Assiened:
Aerospace Medical Research Lab.
State University, AR 72467
A
i
(501) 972-2088
Ik-
Charles Camp
Dss^
PhD
>■
Assistant Professor
Soecialtv:
Civil Engineering
»
Memphis State University
Assiened:
Armament Laboratory
i
Civil Engineering Dept.
Memphis, TN 38152
•V
(901) 678-3169
William Campbell
^SXSSSL
PhD
►
Associate Professor
Soecialtv:
Mathematics
Talladega College
Assiened:
Weapons Laboratory
>•
Math Dept
Talladega, AL 35160
r
y
(205) 362-0206
\
Arnold Carden
Degree:
PhD
V
Professor
Soecialtv:
Metallurgy
r
Alabama, University of
Assiened:
Armament Laboratory
*
PO Box 870278
\
Tuscaloosa, AI. 35487
y
(205) 348-1619
r
V
V
NAME / ADDRESS
DEGREE, SPECIALTY, LABORATORY ASSIGNED
Richard Carlin
Assistant Professor
Alabama, University of
Dept of Chemistry
Tuscaloosa, AL 35487
(205)348-8443 •
Degree;
Snecialtv:
Assigned:
PhD
Chemistry
Frank J. Seiler Research Lab.
Gene Carlisle
Professor
West Texas State University
Dept of Chemistry & Physics
Canyon, ra 79016
(806) 656-2282
Degree:
Snecialtv:
Assigned:
PhD
Inorganic Chemistry
Weapons Laboratory
Chia-Bo Chang
Associate Professor
Texas Tech. Univ.
PO Box 4320
Lubbock, TX 79409
(806) 742-3143
Degree:
Assigned;
PhD
Meteorology
Geophysics Laboratory
Wayne Charlie
Associate Professor
Colorado State University
Dept of Civil Engineering
Fort Collins, CO 80523
(303) 491-5048
Degree;
Snecialtv:
Assigned:
PhD
Civil Engineering
Engineering & Services Center
Chih-Fan Chen
Professor
Boston University
755 Commonwealth Ave.
Boston, MA 02215
(617) 353-2566
Degree;
Snecialtv:
Assigned;
PhD
Engineering
Electronic Systems Division
VI
NAME,/ ADDRESS
DEGREE, SPECIALTY, LABORATORY ASSIGNED
Pyiyuen Chen
Associate Professor
Syracuse University
Dept, of Mathematics
Syracuse, NY 13244
(315) 443-1573
Degree:
Soecialtv:
Assigned:
PhD
Statistics
Human Resources Laboratory
Manpower and Personnel
Muhammad Choudhry
Associate Professor
West Virginia University
PO Box 6101
Morgantown, WV 26506
(304) 293-6375
Degree:
Soecialtv:
Assigned:
PhD
Electrical Engineering
Aero Propulsion Laboratory
Donald Chung
Associate Professor
San Jose State University
Dept of Materials Eng.
San Jose, CA 95192
(408) 924-3873
Degree:
Soecialtv:
Assigned:
PhD
Materials Science
Materials Laboratory
Mingking Chyu
Assistant Professor
Carnegie Mellon University
Dept of Mechanical Eng.
Pittsburgh, PA 15213
(412) 268-3658
Degree:
Soecialtv:
Assigned:
PhD
Mechanical Engineering
Aero Propulsion Laboratory
R. H. Cofer
Associate Professor
Florida Instit. of Tech.
150 W. University Blvd.
Melbourne, FL 32901
(407) 768-8000
Degree:
Soecialtv:
Assigned:
PhD
Electrical Engineering
Avionics Laboratory
NAME / ADDRESS
DEGREE, SPECIALTY, LABORATORY ASSIGNED
William Gofer
Assistant Professor
Washington State University
Dept of Civil & Environ. Eng.
Pullman, WA 99164
(509) 335-3232
Deeree:
Soecialtv:
Assiened:
PhD
Civil Engineering
Weapons Laboratory
John Connolly
Professor
Missouri-Kansas City, Univ. of
Dept of (Chemistry
Kansas City, MO 64110
(816) 276-2286
Deeree:
Soecialtv:
Assiened:
PhD
Chemistry
Materials Laboratory
Gary Craig
Assistant Professor
Syracuse University
Link Hall
DsgSffl,
SBggjal^:
PhD
Electrical Engineering
Rome Air Development Center
Syracuse, NY 13244
(315) 443-4389
■
Donald Dareing
Professor
Florida, University of
237 MEB
Gainesville, FL 3261 1
(904) 392-0827
Deeree:
Soecialtv:
Assiened:
PhD
Mechanical Engineering
Aero Propulsion Laboratory
Vito DelVecchio
Professor
Scranton, University of
Dept of Biology
Scranton, PA 18510
(717) 961-6117
Deeree:
Soecialtv:
Assiened:
PhD
Biochemistry
School of Aerospace Medicine
NAME / ADDRESS
DEGREE, SPECIALTY, LABORATORY ASSIGNED
Paul Deilenback
Assistant Professor
Southern Methodist Univ.
Civil & Mech, Engineering Dept.
DaUas,TX 75275
(214) 692-4172
Degree:
Soecialtv:
Assigned:
PhD
Mechanical Engineering
Aero Propulsion Laboratory
Eustace Deieniak
Associate Professor
Arizona, University of
McKale Ave.
Tucson, AZ 85721
(602) 621-1019
PpfS^'
Soecialtv:
■ Assigned:
PhD
Optics
Amament Laboratory
Janet Dizinno
Assistant Professor
St, Mary’s University
One Camino Santa Maria
San Antonio, TX 78284
(512) 436-3314
Soecialtv:
Assigned:
PhD
Psychology
Wilford Hall Medical Center
Daniel Dolata
Assistant Professor
Arizona, University of
Dept, of Chemistry
Tucson, AZ 85721
(602) 621-6337
Degree:
Soecialtv:
Assigned:
PhD
Chemistry
Frank J. Seiler Research Lab.
Joseph Dreisbach
Professor
Scranton, University of
Chemistry Dept.
Scranton, PA 18510
(717) 961-7519
Degree:
Soecialtv:
Assigned:
PhD
Chemistry
Engineering & Services Center
THIS
PAGE
IS
MISSING
IN
ORIGINAL
DOCUMENT
NAME / ADDRESS
DEGREE, SPECIALTY, LABORATORY ASSIGNED
Dennis Flentge
Associate Professor
CedarviUe College
Box 601
CedarviUe, OH 45314
(513) 766-2211
Degree:
Soecialtv:
Assigned:
PhD
Physical Chemistry
Aero Propulsion Laboratory
Charles Fosha
Associate Professor
Colorado, Univ. of
1867 Austin Bluffs Parkway
Colorado Springs, CO 80918
(719) 548-0602
Pegree:
Soecialtv:
Assigned:
PhD
Electrical Engineering
Armament Laboratory
Lionel Friedman
Professor
Worcester Polytechnic Instit.
100 Institute Rd.
Worcester, MA 01609
(508) 831-5303
Degree:
Soecialtv:
A?signpti;
PhD
Physics
Rome Air Development Center
Daniel FuUer
Department Head
Nicholls State University
Highway 1
Thibodaux, LA 70310
(504) 448-4504
Degree:
Soecialtv:
Assigned:
PhD
Chemistry
Astronautics Laboratory
Ephrahim Garcia
Assistant Professor
New York-Buffalo, State Univ. of
1012 Furnas Hall
Buffalo. NY 14260
(716) 636-3058
Degree:
Soecialtv:
Assigned:
PhD
Aerospace Engineering
Frank J. Seiler Research Lab.
xi
NAME/ ADDRESS
DEGREE, SPECIALTY, LABORATORY ASSIGNED
Daniel Garland
Assistant Professor
Embry-Riddle Aeronautical Univ.
Humanities/Social Sci.
Daytona Beach, FL 32114
(904) 239-6641
Desree:
Soecialtv:
Assiened:
PhD
Psychology
Human Resources Laboratory
Operations Training Division
Thomas Gearhart
Associate Professor
Captial University
Science Hall E. Main St.
Columbus, OH 43209
(614) 236-6800
Degree:
Soecialtv:
Assigned:
PhD
Mathematics
Avionics Laboratory
John George
Professor
Wyoming, University of
Box 3036
Laramie, WY 82071
(307) 766-2383
Degree:
Soecialtv:
Assigned:
PhD
Applied Mathematics
Araament Laboratory
Frederick Gibson
Instructor
Morehouse College
830 Westview Dr. SW
Atlanta, GA 30312
(404) 681-2800
Degree:
Soecialtv:
Assigned:
MS
Applied Mathematics
Armament Laboratory
Ashok Goel
Assistant Professor
Michigan Tech. University
Dept, of Electrical Engineering
Houghton, m 49931
(906) 487-2868
Degree:
Soecialtv:
Assigned:
PhD
Electrical Engineering
Electronic Technology Laboratory
xii
NAME / ADDRESS
DEGREE, SPECIALTY, LABORATORY ASSIGNED
Harold Goldstein
Associate Professor
District of Columbia, Univ. of
4200 Connecticut Ave. N.W.
Washington, DC 20008
(202) 282-7349
Degree:
Sbecialtv:
Assigned:
MS
Transportation Engineering
Human Resources Laboratory
Training Systems
Reinhard Graetzer
Associate Professor
Penn State University
104 Davey Lab.
University Park, PA 16802
(814) 863-0705
Degree:
Specialty:
Assigned:
PhD
Physics
School of Aerospace Medicine
Paul Griffm
Assistant Professor
Georgia Tech.
School of IS YE
Atlanta, GA 30332
(404) 894-2431
Degree:
Specialty:
Assigned:
PhD
Industrial Engineering
School of Aerospace Medicine
William Grissom
Assistant Professor
Morehouse College
830 Westview Dr.
Atlanta, GA 30314
(404) 681-2800
Degree:
Specialty:
Assigned:
MS
Mechanical Engineering
Arnold Engineering Development
Center
David Grossie Deeree: PhD
Assistant Professor Specialty: Chemistry
Wright State University Assigned: Materials Laboratory
Dept of Chemistry
Dayton. OH 45435
(513) 873-2210
NAME / ADDRESS
DEGREE, SPECIALTY, LABORATORY ASSIGNED
Pushpa Gupta
Professor
Maine, University of
321 NeviUe
Orono, ME 04469
(207) 581-3914
Deeree:
Snecialtv:
Assigned;
PhD
Mathematics
School of Aerospace Medicine
Ramesh Gupta
Professor
Maine, University of
Dept of Mathematics
Orono, ME 04469
(207) 581-3913
Deeree:
Soecialtv;
Assigned;
PhD
Mathematical Statistics
School of Aerospace Medicine
Martin Hagan
Associate Professor
Oklahoma State University
School of Elec. & Comp. Sci.
Stillwater, OK 74078
(405) 744-7340
Degree;
Soecialtv;
Assigned:
PhD
Electrical Engineering
Aerospace Medical Research Lab.
Lawrence Hall
Assistant Professor
South Rorida, Univ. of
Dept of Computer Sci.
Tampa, FL 33620
(813) 974-4195
Degree:
Soecialtv;
Assigned:
PhD
Computer Science
Avionics Laboratory
Kevin Hallinan
Assistant Professor
Dayton, Univ. of
Mech. & Aero. Engineering
Dayton, OH 45469
(513) 229-2875
Degree;
Soecialtv;
Assigned;
PhD
Mechanical Engineering
Aero Propulsion Laboratory
XIV
NAME / ADDRESS
DEGREE, SPECIALTY, LABORATORY ASSIGNED
Marvin Hamstad
r Professor
< Denver, Univ. of
Dept of Engineering
Denver, CO 80208
(303) 871-3191
Frances Harackiewicz
^ Assistant Professor
L Southern Illinois Univ.
Technology Bldg. A
^ Carbondale, IL 62901
' (618) 453-7031
Paul Hedman
^ Professor
Brigham Young University
Chemical Engineering Dept
Provo, UT 84602
(801) 378-6238
[Verlin Hinsz
Assistant Professor
North Dakota State Univ.
^ 115 Minatxi Hall
Fargo, ND 58105
^ (701) 237-7082
Chin Hsu
^ Associate Professor
r Washington State Univ.
Y Dept of Elec, and Comp. Eng.
, Pullman, WA 99164
(509) 335-2342
V-
V-
*
Degree:
Specialty:
Assigned:
Degree:
Specialty:
Assigned:
Degree:
Specialty:
Assigned:
Degree:
Specialty:
Assigned:
Degree:
Specialty:
Assigned:
PhD
Solid Mechanics
Flight Dynamics Laboratory
PhD
Electrical Engineering
Rome Air Development Center
PhD
Chemical Engineering
Aero Propulsion Laboratory
PhD
Psychology
Human Resources Laboratory
Logistics & Human Factors
PhD
Electrical Engineering
Flight Dynamics Laboratory
XV
NAME / ADDRESS
DEGREE, SPECIALTY, LABORATORY ASSIGNED
Ming-Shu Hsu
Associate Professor
Portland, Univ. of
5000 N. Willamette Blvd.
Portland, OR 97203
(503) 283-7436
Decree:
Soecialtv:
Assicned:
PhD
Mechanical Engineering
Flight Dynamics Laboratory
Delayne Hudspeth
Associate Professor
Texas-Austin, Univ. of
College of Education
Austin, TX 78712
(512) 471-5211
Decree:
Soecialtv:
Assicned:
PhD
Education
Human Resources Laboratory
Manpower & Personnel Div.
Manuel Huerta
Professor
Miami, Univ. of
PO Box 248046
Coral Gables, FL 33124
(305) 284-2323
DSggg:
Soecialtv:
Assicned:
PhD
Physics
Armament Laboratory
David Hui
Associate Professor
New Orleans, Univ. of
Dept of Mech. Engineering
New Orleans, LA 70148
(504) 286-6192
Decree:
Soecialtv:
Assicned:
PhD
Aerospace Engineering
Flight Dynamics Laboratory
George Jumper
Associate Professor
Worcester Poly. Instit.
100 Institute Rd.
Worcester, MA 01609
(508) 831-5368
Decree:
Soecialtv:
Assicned:
PhD
Mechanical Engineering
Geophysics Laboratory
XVI
NAME / ADDRESS
DEGREE, SPECIALTY, LABORATORY ASSIGNED
Prasad Kadaba
Professor
Kentucky, Uaiv. of
Rm. 453 Andereon Hall
Lexington, k f 40506
W) 257-2943
lee: PhD
ialtv: Physics
gned; Materials Laboratory
Ngozi Kainalu
Lecturer
Ctdifomia Poly. Univ.
Mechancial Engineering Dept.
San Luis Obispo, CA 9340'/
(805) 756-1336
Specialty; Mechanical Engineering
As.sighed: Frank J. Seiler Research Lab.
Gillray Kandel
Professor
Rensselaer Poly. Instit.
8th St.
Troy, NY 12180
(518) 276-8269
Assigned;
PhD
Experimental P.<!ychology
Human Resources Laboratory
Operations Training bivision
Mohammad Karim
Associate Professor
Dayton, University of
300 College Park
Dayton, OH 45469
(513) 229-3611
Degree;
Specialty;
Assigned;
PhD
Electrical Engineering
Avionics Laboratory
Siavash Kassemi
Assistant Professor
Colorado, University of
1867 Austin Bluffs Parkway
Colorado Springs, CO 80918
(719) 593-3326
Assigned;
PhD
Aerospace Engineering
Frank J. Seiler Research Lab.
NAME / ADDRESS
DEGREE, SPECIALTY, LABORATORY ASSIGNED
Yulian Kin
Associate Professor
Purdue Calumet
Engineering Dept
Hammond, IN 46323
(219) 989-2684
Degree:
Soecialtv:
Assigned:
PhD
Mechanical Engineering
Flight Dynamics Laboratory
Kevin Kirby
Assistant Professor
Wright State University
3171 Research Blvd.
Kettering, OH 45420
(513) 259-1373
Degree:
Soecialtv:
Assigned:
PhD
Computer Science
Avionics Laboratory
David Kirkner
Associate Professor
Notre Dame, Univ, of
Dept of Civil Engineering
Notre Dame, IN 46556
(219) 239-6518
Degree:
Soecialtv:
Assigned:
PhD
Solid Mechanics
Engineering & Services Center
Ashok Krishnamurthy
Assistant Professor
Ohio State University
2015 Neil Ave.
Columbus, OH 43210
(614) 292-5604
Degree:
Soecialtv:
Assigned:
PhD
Electrical Engineering
Aerospace Medical Research Lab.
Paul Kromann
Associate Professor
Fort Valley State College
Campus Box 4821
Fort Valley, GA 31030
(912) 825-6245
Degree:
Soecialtv:
Assigned:
PhD
Chemistry
Engineering & Services Center
xviii
NAME /ADDRESS
DEGREE, SPECIALTY, LABORATORY ASSIGNED
Jeffrey Kuhn
Associate Professor
Michigan State University
309 Physics-Astronomy Bldg.
East Lansing, MI 48824
(517) 353-2986
Deeree:
Soecialtv:
Assiened;
PhD
Physics
Geophysics Laboratory
Kyung Kwon
Associate Professor
Tuskege University
Chemistry Dept.
Tuskegee, AL 36088
(205) 727-8089
Degree;
Soecialtv:
Assiened:
PhD
Chemical Engineering
Engineering & Services Center
Joseph Lambert
Professor
Northwestern University
2145 Sheridan Rd.
Evanston, IL 60208
(708) 491-5437
Deeree;
Soecialtv:
Assiened:
PhD
Chemistry
Materials Laboratory
Gary Leatherman
Assistant Professor
Worcester Polytechnic Instit.
100 Institute Rd.
Worcester, MA 01609
(508) 831-5229
Deeree:
Soecialtv;
Assiened;
PhD
Materials Science
Materials Laboratory
Byung-Lip Lee
Associate Professor
Pennsylvania State University
227 Hammond Bldg.
University Park, PA 16802
(814) 865-7829
Deeree;
Soecialtv;
Assiened:
PhD
Materials Science
Flight Dynamics Laboratory
xix
NAME / ADDRESS
DEGREE, SPECIALTY, LABORATORY ASSIGNED
Tzesan
Associate Professor
Western Illinois Univ.
900 W. Adams St.
Macomb, IL 61455
(309) 298-1485
Desree:
Soecialtv:
Assigned:
PhD
Applied Mathematics
Aerospace Medical Research Lab.
Won-Kyoo Lee
Associate Professor
Ohio State University
140 W. 19th Ave.
Columbus, OH 43211
(614) 292-6605
Degree:
Soecialtv:
Assigned:
PhD
Chemical Engineering
Materials Laboratory
Paul Lemke
Professor
Auburn University
131 Funchess Hail
Auburn University, AL 36849
(205) 844-1662
Dgffgp;,
Soecialtv:
Assigned:
PhD
Molecular Biology
School of Aerospace Medicine
Sigmund Lephait
Lecturer
Melbourne Univ. Australia
Parkville 3052
Victoria Australia,
(03) 344-5158
Dffgee:
Specialty:
Assigned:
PhD
Biomechanics
Aerospace Medical Research Lab.
Shannon Lieb
Associate Professor
Butler University
4600 Sunset Ave.
Indianapolis, IN 46208
(317) 283-9410
Degree:
Soecialtv:
Assigned:
PhD
Physical Chemistry
Astronautics Laboratory
XX
NAME / ADDRESS
DEGREE, SPECIALTY, LABORATORY ASSIGNED
Hao Ling
Assistant Professor
Texas-Austin, Univ. of
Dept of Elec. & Comp. Eng.
Austin, TX 78712
(512) 471-1710
Degree:
Soecialtv:
Assigned:
PhD
Electrical Engineering
Rome Air Development Center
C. Randal Lishawa
Assistant Professor
Utica College
Buifstone Rd.
Utica, NY 13502
(315) 792-3139
Degree:
Soecialtv:
Assigned:
PhD
Physical Chemistry
Geophysics Laboratory
Vernon Matzen
Associate Professor
North Carolina State Univ.
Box 7908
Raleigh, NC 27695
(919) 737-2331
DtsssSi
Soecialtv:
Assigned:
PhD
Structural Mechanics
Flight Dynamics Laboratory
Michael McFarland
Assistant Professor
Utah State Univ.
Utah Water Research Lab.
Logan, UT 84322
(801) 750-3196
Degree:
Soecialtv:
Assigned:
PhD
Biological Engineering
Engineering & Services Center
Perry McNeill
Professor
North Texas, Univ. of
PO Box 13198
Denton, TX 76203
(817) 565-2846
Degree:
Soc :ialtv:
As.‘ igned:
PhD
Education
Engineering & Services Center
XXI
NAME / ADDRESS
DEGREE, SPECIALTY, LABORATORY ASSIGNED
Miguel Medina
Associate Professor
Duke University
Dept of Civil Engineering
Durham, NC 27706
(919) 660-5195
Deeree:
Soecialtv:
Assigned:
PhD
Water Resources
Occupational & Environmental
Health Laboratory
Richard Miers
Associate Professor
Indiana Univ. - Purdue Univ.
2101 Coliseum Blvd. E.
Fort Wayne, IN 46805
(219) 481-6154
Degree:
Soecialtv:
Assigned:
PhD
Physics
Avionics Laboratory
William Moor
Associate Professor
Arizona State Univ.
College of Engineering
Tempe, AZ 85287
(602) 965-4022
Degree:
Soecialtv:
Assigned:
PhD
Industrial Engineering
Human Resources Laboratory
Operations Training Division
Carlyle Moore
Associate Professor
Morehouse College
830 Westview Dr.
AUanta,GA 30314
(404) 681-2800
Degree:
Soecialtv:
Assigned:
PhD
Physics
Arnold Engineering Development
Center
Kevin Moore
Assistant Professor
Idaho State Univ.
Box 8060
Pocatello, ID 83209
(208) 236-4188
Degree:
Soecialtv:
Assigned:
PhD
Electrical Engineering
Armament Laboratory
xxii
NAME / ADDRESS
DEGREE, SPECIALTY, LABORATORY ASSIGNED
Rex Moyer
Professor
Trinity University
715 Stadium Dr.
San Antonio, TX 78212
(512) 736-7242
Decree:
Soecialtv:
Assigned:
PhD
Microbiology
School of Aerospace Medicine
Arnold Nelson
Assistant Professor
Louisiana State Univ.
112 Long Field House
Baton Rouge, LA 70803
(504) 388-3114
Dggrpe:
Soecialtv:
Assigned:
MS
Physical Education
School of Aerospace Medicine
Kirk Noidyke
Instructor
Xavier University
Dept, of Biology
New Orleans, LA 70125
(504) 483-7527
Pggree:
Soecialtv:
Assigned:
MS
Zoology
Occupational & Environmental
HealUi Laboratory
Olin Norton
Researcher
Mississippi State Univ.
PO Drawer MM
Mississippi State, MS 39762
(601) 325-2105
Degree:
Assigned:
PhD
Mechanical Engineering
Arnold Engineering Development
Center
Muhammad Numan
Assistant Professor
Indiana Univ. of Pennsylvania
45 Weyandt HaU
Indiana, PA 15705
(412) 357-2318
Degree:
Soecialtv:
Assigned:
PhD
Physics
Electronic Technology Laboratory
xxm
NAME/ ADDRESS
DEGREE, SPECIALTY, LABORATORY ASSIGNED
Amit Patra
Associate Professor
Puerto Rico, Univ. of
PO Box 5000
De^e:
Soecialtv:
Assigned:
PhD
Mechanical Engineering
Aerospace Medical Research Lab.
Mayaguez, PR 00709
(809) 832-4040
Shietung Peng
Assistant Professor
Maryland-Baltimore, Univ. of
5401 Wilkens Ave.
Baltimore, MD 21228
(301) 455-3540
Degree:
Soecialtv:
Assigned:
PhD
Computer Science
Rome Air Development Center
Richard Peters
Assistant Professor
Vanderbilt University
Box 6091 Station B
Nashville, TN 37235
(615) 322-7924
Degree:
Soecialtv:
Assigned:
PhD
Electrical Engineering
Arnold Engineering Development
Center
Bernard Piersma
Professor
Houghton College
Dept, of Chemistry
Houghton, NY 14744
(716) 567-9301
Degree:
Soecialtv:
Assigned:
PhD
Physical Chemistry
Frank J. Seiler Research Lab.
Thomas Pollock
Associate Professor
Texas A«feM University
Dept, of Aerospace Engineering
College Station, TX 77843
(409) 845-1686
Degree:
Soecialtv:
Assigned:
PhD
Materials Science
Astronautics Laboratory
xxiv
NAME / ADDRESS
DEGREE, SPECIALTY, LABORATORY ASSIGNED
Thomas Posbergh
Assistant Professor
Minnesota, Univ. of
107 Akerman Hall
Minneapolis, MN 55455
(612) 625-2871
Degree:
Soecialtv:
Assiened:
PhD
Electrical Engineering
Frank J. Seiler Research Lab.
James Price
Professor
Iowa, University of
W140 Seashore Hall
Iowa City, lA 52242
(319) 335-2497
Deerce:
Soecialtv:
Assiened:
PhD
Sociology
Human Resources Laboratory
Manpower & Personnel Div.
Gandikota Rao
Professor
St. Louis University
3507 Laclede Ave.
St Louis, MO 63103
(314) 658-3115
Deerce:
Soecialtv:
Assiened:
PhD
Meteorology
Geophysics Laboratory
K. Sankara Rao
Professor
North Dakota State Univ.
Dept of Electrical Engineering
Fargo, ND 58105
(701) 237-7217
Degree:
Soecialtv:
Assiened:
PhD
Electrical Engineering
Aero Propulsion Laboratory
Craig Rasmussen
Assistant Professor
Utah State University
CASS UMC 4405
Logan, UT 84322
(801) 750-2967
Deeree:
Soecialtv:
Assiened:
PhD
Physics
Geophysics Laboratory
XXV
NAME / ADDRESS
DEGREE, SPECIALTY, LABORATORY ASSIGNED
Joan Rentsch
Assistant Professor
Wright State University
Dept, of Psychology
Dayton, OH 45435
(513) 873-2218
Desree:
Soecialtv:
Assigned:
PhD
Industrial Psychology
Human Resources Laboratory
Logistics & Human Factors
Michael Resch
Assistant Professor
Nebraska-Lincoln, Univ. of
212 Bancroft Hall
Lincoln, NE 68588
(402) 472-2354
Decree:
Soecialtv:
Assigned:
PhD
Materials Science
Materials Laboratory
Donald Robinson
Assistant Professor
Xavier University
7325 Palmetto St.
New Orleans, LA 70125
(504) 483-7371
Degree:
So^ialtv:
Assigned:
PhD
Chemistry
School of Aerospace Medicine
Larry Roe
Assistant Professor
Virginia Poly. Instit. State Univ.
Mechanical Engineering Dept.
Blacksburg, VA 24061
(703) 231-7295
Degree:
Specialty:
Assigned:
PhD
Mechanical Engineering
Aero Propulsion Laboratory
John Russell
Associate Professor
Florida Inst, of Tech.
150 W. University Blvd.
Melbourne, FL 32901
(407) 768-8000
Degree:
Soecialtv:
Assigned:
PhD
Aerospace Engineering
Arnold Engineering Development
Center
xxvi
NAME /ADDRESS
DEGREE, SPECIALTY, LABORATORY ASSIGNED
Daniel Ryder
Assistant Professor
Tufts University
Chemical Engineering Dept.
Medford, MA 02155
(617) 381-3446
Degree:
Soecialtv:
Assigned:
PhD
Chemical Engineering
Rome Air Development Center
John Scharf
Chairman
Carroll College
Dept of Math.
Helena, MT 59625
(406) 442-3450
Degree:
Soecialn':
Assigned:
MS
Civil Engineering
Engineering & Services Center
Johanna Schruben
Associate Professor
Houston-Victoria, Univ. of
2302C Red River
Victoria. TX 77901
(512) 576-3151
Degree:
Specialty;
Assigned:
PhD
Mathematics
Weapons Laboratory
Martin Schwartz
Professor
North Texas, Univ. of
PO Box 5068
Denton, TX 76203
(817) 565-3524
Degree;
Specialty:
Assigned;
PhD
Physical Chemistry
Materials Laboratory
David Senseman
Professor
Texas-San Antonio, Univ. of
Div. of Life Sciences
San Antonio, TX 78285
(512) 691-5485
Degree:
Soecialtv:
Assigned:
PhD
Biology
School of Aerospace Medicine
xxvii
NAME / ADDRESS
DEGREE, SPECIALTY, LABORATORY ASSIGNED
Brian Shelburne
Associate Professor
Wittenberg University
Box 720
Springfield, OH 45501
(513) 327-7862
Desree:
Sbecialtv:
Assigned;
PhD
Mathematics
Avionics Laboratory
Behiooz Shirazi
Assistant Professor
Southern Methodist University
Dept of Comp, Sci, & Engineering
Dallas, TX 75275
(214) 692-2874
Deeree:
Soecialtv;
Assiened:
PhD
Computer Science
Rome Air Development Center
Leonard Shyles
Associate I^fessor
Villanova University
Dept of Communication Arts
Vmanova,PA 19085
(215) 645-7923
Deeree:
Soecialtv:
Assiened:
PhD
Communication
Aerospace Medical Research Lab.
William Siuru
Associate Professor
Colorado, Univ. of
1867 Austin Bluffs Parkway
Colorado Springs, CO 80918
(719) 548-0602
Deeree:
Soecialtv:
Assiened:
PhD
Mechanical Engineering
Armament Laboratory
Eleanor Smith
Assistant Professor
Florida A&M University
406 Perry-Paige
Tallahassee, FL 32307
(904) 599-3821
Deeree:
Soecialtv:
Assiened:
PliD
Sociology
Human Resources Laboratory
Training Systems Division
xxviii
NAME / ADDRESS
DEGREE, SPECIALTY, LABORATORY ASSIGNED
Wayne Smith
Associate Professor
Mississippi State University
Drawer CS
Mississippi State, MS 39762
(601) 325-2642
Dearee:
Soecialtv:
Assiened:
PhD
Computer Science
Rome Air Development Center
Kenneth Sobel
Associate Professor
New York, City College of
Dept of Electrical Engineering
New York, NY 10031
(212) 690-4241
Desree:
Soecialtv:
Assigned:
PhD
Electrical Engineering
Armament Laboratory
Glenn Stark
Assistant Professor
Wellesley College
Dept of Physics
WeUesley,MA 02181
(617) 235-0320
Deerce:
Soecialtv:
Assiened:
PhD
Physics
Geophysics Laboratory
Stanley Stephenson
Associate I^fessor
Southwest Texas State Univ.
CIS/ADS
San Marcos, TX 78666
(512) 245-2291
Deeree:
Soecialtv:
Assiened:
PhD
Psychology
Human Resources Laboratory
Training Systems Division
Chun Su
Assistant Professor
Mississippi State Univ.
Dept of Physics
Mississippi State, MS 39762
Deeree:
Soecialtv:
Assiened:
PhD
Physics
Arnold Engineering Development
Center
(601) 325-2931
xxix
NAME / ADDRESS
DEGREE, SPECIALTY, LABORATORY ASSIGNED
Richard Swope
Professor
Trinity University
715 Stadium Dr.
San Antonio, TX 78212
(512) 736-7514
Degree:
Soecialtv:
Assigned:
PhD
Mechanical Engineering
School of Aerospace Medicine
John Szarek
Assistant Professor
Marshall University
1542 Spring Valley Dr.
Huntington, WV 25755
(304) 696-7314
Degree:
Soecialtv:
Assigned:
PhD
Pharmaceutical
School of Aerospace Medicine
Kaveh Tagavi
Associate Professor
Kentucky, Univ. of
242 Anderson Hall
Lexington, KY 40506
(606) 257-2739
Degree:
Soecialtv:
Assigned:
PhD
Mechanical Engineering
Aero Propulsion Laboratory
Devki Talwar
Assistant Professor
Indiana Univ. of Pennsylvania
Dept of Physics
Indiana, PA 15705
(412) 357-4589
Pegree:
Soecialtv:
Assigned:
PhD
Physics
Electronic Technology Laboratory
Richard Tankin
Professor
Northwestern University
Dept, of Mechanical Engineering
Evanston, IL 60208
(708) 491-3532
Degree:
Soecialtv:
Assigned:
PhD
Engineering
Aero Propulsion Laboratory
XXX
NAME / ADDRESS
DEGREE, SPEQALTY, LABORATORY ASSIGNED
Roger Thompson
Assistant Professor
Pennsylvania State Univ.
233 Hammond Bldg.
University Park, PA 16802
(814) 863-0968
Degree:
Soecialtv:
Assigned:
PhD
Mechanical Engineering
Astronautics Laboratory
Steven Trogdon
Associate Professor
Minnesota, Univ, of
108 Heller Hall
Duluth, MN 55812
(218) 726-6173
Degree:
Soecialtv:
Assigned:
PhD
Mechanics
Armament Laboratory
Hai-Lung Tsai
Assistant Professor
Missouri-Rolla, Univ. of
DepL of Mech. & Aero.
RoUa,MO 65401
(314) 341-4945
Degree:
Segcialty;
Assigned:
PhD
Mechanical Engineering
Materials Laboratory
Pamela Tsang
Assistant Professor
Wright State University
309 Oelman
Dayton, OH 45435
(513) 258-2687
Degree:
Soecialtv:
Assigned:
PhD
Engineering Psychology
Human Resources Laboratory
Logistics & Human Factors
Ronald VanEtten
Associate Professor
Illinois State University
500 W. Gregory
Normal, IL 61761
(309) 438-8346
Degree:
Soecialtv:
Assigned:
MS
Computer Science
Rome Air Development Center
XXXI
NAME / ADDRESS
DEGREE, SPECIALTY, LABORATORY ASSIGNED
George Veyera
Assistant Ftofessor
Rhode Island, Univ. of
Dept of Qvil Engineering
Kingston, RI 02881
(401) 792-2684
Degree:
Soecialtv:
Assigned:
PhD
Civil Engineering
Engineering & Services Center
Hung Vu
Assistant Professor
California State Univ.
1250 Bellflower Blvd.
Long Beach, CA 90840
(213) 985-1524
Dtsnsk
Soecialtv:
Assigned:
PhD
Applied Mechanics
Astronautics Laboratory
Bonnie Walker
Assistant Professor
Central State University
Psychology Dept
Wilberforce, OH 45384
(513) 376-6516
Degree:
Soecialtv:
Assigned:
PhD
Experimental Psychology
Aerospace Medical Research Lab.
Steven Waller
Associate Professor
South Dakota, Univ. of
Dept of Physiol. & Pharmacol.
Vermillion, SD 57069
(605) 677-5157
Desssi
Soecialtv:
Assigned:
PhD
Pharmacology
School of Aerospace Medicine
Peter Walsh
Professor
Fairleigh Dickinson Univ.
Dept of Electrical Engineering
Teaneck, NJ 07666
(201) 692-2493
Degree:
Soecialtv:
Assigned:
PhD
Physics
Weapons Laboratory
xxxii
NAME / ADDRESS
DEGREE, SPECIALTY, LABORATORY ASSIGNED
Lorin Weber
Professor
Ricks College
Physics Dept.
Rexburg, ID 83460
(208) 356-1907
Desree:
Soecialtv:
Assigned;
MS
Science Education
Occupational & Environmental
Health Laboratory
Kevin Whitaker
Assistant Professor
Alabama, University of
Box 870280
Tuscaloosa, AL 35487
(205) 348-7366
Deeiee:
Soecialtv:
Assiened:
PhD
Aerospace Engineering
Arnold Engineering Development
Center
Trevor Williams
Assistant Professor
Cincinnati, University of
ML 70
Cincinnati, OK 45221
(513) 556-3221
Deeree:
Soecialtv:
Assiened:
PhD
Control Theory
Astronautics laboratory
John Wills
Professor
Indiana University
Physics Dept.
Bloomington, IN 47405
(812) 855-1479
Deeice:
Specialty:
Assiened:
PhD
Physics
Geophysics Laboratory
Martin Wilner
Professor
Lowell, University of
1 University Ave.
Lowell, MA 01854
(508) 934-3786
Degree:
Soecialtv:
Assigned:
PhD
Physics
Rome Air Development Center
XXXlll
NAME / ADDRESS
DEGREE, SPECIALTY, LABORATORY ASSIGNED
William Wolfe
Associate Professor
Ohio State University
470 Hitchcock Hall
Columbus, OH 43210
(614) 292-0790
Deeree:
Soecialtv:
Assigned:
PhD
Engineering
Flight Dynamics Laboratory
James Wolper
Assistant Ptofessor
Hamilton College
Dept, of Math & Comp. Sci.
Clinton, NY 13323
(315) 859-4417
Degree;
Soecialtv:
Assigned;
PhD
Mathematics
Rome Air Development Center
Hsien-Yang Yeh
Associate Professor
California State Univ.
1250 Bellflower Blvd.
Long Beach, CA 90840
(213) 985-4611
Degree:
Soecialtv:
PhD
Structural Mechanics
Astronautics Laboratory
Lawrence Zavodney
Assistant Professor
Ohio State Univ.
209 Boyd Laboratoiy
Columbus, OH 43210
(614) 292-2209
Degree;
Soecialtv:
Assigned;
PhD
Mechanical Engineering
Flight Dynamics Laboratory
Wayne Zimmermann
Associate Professor
Texas Woman’s University
PO Box 22865
Denton, TX 76204
(817) 898-2166
Degree;
Soecialtv;
Assigned:
PhD
Applied Mathematics
Weapons Laboratory
XXXIV
PARTICIPANT LABORATORY ASSIGNMENT
XXXV
C. PARTICIPANT LABORATORY ASSIGNMENT (Page 1)
1990 USAF/UES SUMMER FACULTY RESEARCH PROGRAM
AERO PROPULSION LABORATORY (WRDC/APL)
(Wright-Patterson Air Force Base)
1.
Muhammad Choudhry
7.
Paul Hedman
2.
Mingking Chyu
8.
K. Sankara Rao
3.
Donald Dareing
9.
Larry Roe
4.
Paul Dellenback
10.
Kaveh Tagavi
5.
Dennis Eentge
11.
Richard Tankin
6.
Kevin Hallinan
ARMAMENT LABORATORY (ATL)
(Eglin Air Force Base)
1.
Charles Camp
7.
Manuel Huerta
2.
Arnold Carden
8.
Kevin Moore
3.
Eustace Dereniak
9.
William Siuru
4.
Charles Fosha
10.
Kenneth Sobel
5.
John George
11.
Steven Trogdon
6.
Frederick Gibson
HARRY G.
ARMSTRONG AEROSPACE MEDICAL RESEARCH LABORATORY (AAMRL)
(Wright-Patterson AFB)
1.
Richard Backs
6.
Tzesan Lee
2.
Larry Byrd
7.
Sigmund Lephart
3.
John Duncan
8.
AmitPatra
4.
Martin Hagan
9.
Leonard Shyles
5.
Ashok Kiishnamurthy
10.
Bonnie Walker
ARNOLD ENGINEERING DEVELOPMENT CENTER
1
n
(Arnold Air Force Base)
1.
William Grissom
5.
John Russell
2.
Carlyle Moore
6.
Chun Su
3.
Olin Norton
7.
Kevin Whitaker
4.
Richard Peters
ASTRONAUTICS LABORATORY (AL)
Edwards Air Force Base)
1. Daniel Fuller
2. Shannon Lieb
3. Thomas Pollock
4. Roger Thompson
5. Hung Vu
6. Trevor Williams
7. Hsien-Yang Yeh
xxxvi
C. PARTICIPANT LABORATORY ASSIGNMENT (Page 2'
AVIONICS LABORATORY (WRDC/AL)
(Wright-Patterson Air Force Base)
1. Thomas Abraham
2. Michael Breen
3. R. H. Gofer
4. Thomas Gearhart
5. Lawrence Hall
ELECTRONIC SYSTEMS DIVISION (ESD)
(Hanscom Air Force Base)
1. Chih-Fan Chen
ELECTRONIC TECHNOLOGY LABORATORY (WRDC/ETL)
(Wright-Patterson Air Force Base)
1. Ashok Goel
2. Muhammad Numan
3. Devki Talwar
ENGINEERING AND SERVICES CENTER (ESC)
(Tyndall Air Force Base)
1. William Bannister 7.
2. Mark Bmsseau 8.
3. Wayne Charlie 9.
4. Joseph Dreisbach 10.
5. David Kirkner 11.
6. Paul Kromann
Kyung Kwon
Michael McFarland
Perry McNeill
John Scharf
George Veyera
6. Mohammad Karim
7. Kevin Kirby
8. Richard Miers
9. Brian Shelburne
FLIGHT DYNAMICS LABORATORY (WRDC/FDL)
(Wright-Patterson Air Force Base)
1. John Bay 7.
2. Franklin Eastep 8.
3. Marvin Hamst^ 9.
4. Chin Hsu 10.
5. Ming-Shu Hsu 11.
6. David Hui
Yulian Kin
Byung-Lip Lee
Vernon Matzen
William Wolfe
Lawrence Zavodney
FRANK J. SEILER RESEARCH LABORATORY (FJSRL)
(USAF Academy)
1. Theodore Burkey 5. Ngozi Kamalu
2. Richard Carlin 6. Siavash Kassemi
3. Daniel Dolata 7, Bernard Piersma
4. Ephrahim Garcia 8. Thomas Posbergh
xxxvu
C. PARTICIPANT LABORATORY -ASSIGNMENT (Page 31
GEOPHYSICS LABORATORY (AFGL)
(Hanscom Air Force Base)
1.
Theodore Aufdemberge
7.
C. Randal Lishawa
2.
Frank Battles
8.
Gandikota Rao
3.
Reuben Benumof
9.
Craig Rasmussen
4.
Chia-Bo Chang
10.
Glenn Stark
5.
George Jumper
11.
John Wills
6.
Jeffrey Kuhn
HUMAN RESOURCES LABORATORY (HRL)
(Brooks, Williams, and Wright-Patterson Air Force Bases)
1.
Margaret Batschelet
8.
Gillray Kandel
2.
Pinyuen Chen
9.
William Moor
3.
James Dykes
10.
James Price
4.
Daniel Garland
11.
Joan Rentsch
5,
Harold Goldstein
12.
Eleanor Smith
6.
Verlin Hinsz
13.
Stanley Stephenson
7.
Delayne Hudspeth
14.
Pamela Tsang
MATERIALS LABORATORY (ML)
(Wright-Patterson Air Force Base)
1.
Donald Chung
7.
Gary Leatherman
2.
John Connolly
8.
Won-Kyoo Lee
3.
Sheiif Elwakil
9.
Michael Resch
4.
David Grossie
10.
Martin Schwartz
5.
Prasad Kadaba
11.
Hai-Lung Tsai
6.
Joseph Lambert
OCCUPATIONAL AND ENVIRONMENTAL HEALTH LABORATORY (OEHL)
(Brooks Air Force Base)
1.
David Buckalew
3.
Kirk Nordyke
2.
Miguel Medina
4.
Lorin Weber
ROME AIR DEVELOPMENT CENTER (RADC)
(Griffiss Air Force Base)
1.
Charles Alajajian
7.
Daniel Ryder
2.
Gary Craig
8.
Behrooz Shirazi
3.
Lionel Friedman
9.
Wayne Smith
4.
Frances Harackiewicz
10.
Ronald VanEtten
5.
Hao Ling
11.
Martin Wilner
6.
Shietung Peng
12.
James Wolper
xxxvm
C. PARTICIPANT LABORATORY ASSIGNMENT (Page 4)
SCHOOL OF AEROSPACE MEDICINE (SAM)
(Brooks Air Force Base)
1.
Phillip Bishop
10.
Paul Lemke
2.
Robert Blystone
11.
Rex Moyer
3.
Bruno Breitmeyer
12.
Arnold Nelson
4.
Vito DelVecchio
13.
Donald Robinson
5.
Randall Dupre
14.
David Senseman
6.
Reinhard Graetzer
15.
Richard Swope
7.
Paul Griffin
16.
John Szarek
8.
Pushpa Gupta
17.
Steven Waller
9.
Ramesh Gupta
WEAPONS LABORATORY (WL)
(Kirtland Air Force Base)
1.
William Campbell
4.
Johanna Schruben
2.
Ggene Carlisle
5.
Peter Walsh
3.
William Cofer
6.
Wayne Zimmerman
WILFORD HALL MEDICAL CENTER (WHMC)
(Lackland Air Force Base)
1. Janet Dizinno
xxxix
RESEARCH REPORTS
xl
RESEARCH REPORTS
1990 SUMMER FACULTY RESEARCH PROGRAM
Technical
Report
Number
Title
Professor
Volume I
Armament Laboratory
1
Simple Models for Predicting Runway
Failure Due to Blast Loading
Dr. Charles Camp
2
Physical Aspects of the Penetradon of
Reinforced Concrete Slabs
Dr. Arnold Carden
3
Solid-State Imager Replacement for a
High-Speed Film Camera
Dr. Eustace Dereniak
4
Evaluation of Weapon Target Allocation
Algorithms
Dr. Charles Fosha
5
Methods Which Accelerate Convergence in
Iterative CFD Solvers
Dr. John George
6
Designing a Binary Phase Only Filter Via
the Genetic Algorithm
Mr. Frederick Gibson
7
Two-Dimensional Simulation of Railgun
Plasma Armatures
Dr. Manuel Huerta
8
Neural Networks for Guidance, Navigation,
and Control of Exoatmospheric Interceptors
*** Not Publishable at this Time ***
Dr. Kevin Moore
9
Gunner Performance in the BSTING
Fire Control System
Dr. William Siuru
10
Robust Eigenstructure Assignment with
Application to Missile Control
Dr. Kenneth Sobel
11
Multiple Point Detonation Modeling
Dr. Steven Trogdon
xli
Volume I
Arnold Engineering Development Center
12
Development of a Combustion Model for
Liquid Film Cooled Rocket Engines
Mr. William Grissom
13
Feasibility of Measuring Pulsed X-Ray
Spectra Using Photoactivadon of
Nuclear Isomers
Dr. Carlyle Moore
14
Combustion of Carbon Particles in the Plume
of a Eare
Dr. Olin Nonon
15
Noise Reduction in Rocket Test Videos using
Mathematical Morphology
Dr. Richard Peters
16
On the Hazard of Combustion Chamber
Oscillations in a Large Freejet Test Cell
Dr. John Russell
17
Laser-Induced Euorescence of Nitric Oxide
Dr. Chun Su
18
An Algorithm for Defining the Shape of a
Plume Exhausting from a Rectangular Nozzle
Dr. Kevin Whitaker
Astronautics Laboratory
19
Strain Enhancing Binder Blends for Use
in Rocket Propellants
Dr. Daniel Fuller
20
A Development of Molecular Modeling
Techniques to Study Intermolecular Forces
Found Between Solid Rocket Oxidizers
and Their Binding Agents
Dr. Shannon Lieb
21
Design and Analysis of Reaction Wheel
Actuators for ASTREX
Dr. Thomas Pollock
22
Component Design for the Multi-Body
Dynamics Experiment
Dr. Roger Thompson
23
Control Design of ASTREX Test Article
Dr. Hung Vu
24
Identification and Control of Hexible
Spacecraft
Dr. Trevor Williams
xlii
Volume I
25 Investigating the Loading Rate Effect on
the Crack Growth Behavior in a Composite
Solid Propellant
Electronics Systems Division
26 Walsh Function Analysis of Impulse Radar
Engineering and Services Center
27 Kigh Oxygen/Carbon Ratio Fuel Candidates for
Clean Air Fire Fighting Facility Project
28 Rate-Limited Mass Transfer and Solute
Transport
29 Centrifuge Modeling of Explosive Induced
Stress Waves in Unsaturated Sand
30 Pathways of 4-Nitrophenol Degradation
31 Permanent Deformations in Airfield Pavement
Systems with Thick Granular Layers
32 The High-Speed Separation and Identification
of Jet Fuel
33 Utilization of Ion Exchange Resins for
the Purification of Plating Baths
34 Methanotrophic Cometabolism of
Trichloroethylene (TCE) in a Two Stage
Bioreactor System
35 Submicron Antennas for Solar Energy
Conversion
36 Dynamic Analysis of Impulse Loading on
Laminated Composite Plates Using
Normal-Mode Techniques
37 A Specimen Preparation Technique for
Microsmictral Analysis of Unsaturated Soil
Dr. Hsien-Yang Yeh
Dr. Chih-Fan Chen
Dr. William Bannister
Dr. Mark Brusseau
Dr. Wayne Charlie
Dr. Joseph Dreisbach
Dr. David Kirkner
Dr. Paul Kromann
Dr. Kyung Kwon
Dr. Michael McFarland
Dr. Perry McNeill
Mr. John Scharf
Dr. George Veyera
xliii
Volume n
Frank J. Seiler Research Laboratory
38
Thermal Decomposition of NTO and
NTO/TNT Mixtures
Dr. Theodore Burkey
39
Transition Metal Carbonyl Complexes in
Ambient-Temperature Molten Salts and
Alkali Metal Reductions at Tungsten and
Mercury Filin Electrodes in Entered Neutral
Aluminum Chloride: l-Methyl-3-Ethylimidazolium
Chloride Molten Salts
Dr. Richard Carlin
40
Expert Guide: Using Artificial Intelligence
Techniques to Help Chemists Utilize Numerical
Programs
Dr. Daniel Dolata
41
Control of a Complex Flexible Structure
Utilizing Space-Realizable Linear Reaction
Mass Actuators
Dr. Ephrahim Garcia
42
Particle Image Displacement Velocimetry
(PIDV) Measurements in Dynamic Stall
Phenomena
Dr. Ngozi Kamalu
43
A Preliminary Analysis of Symbolic
Computational Technique for Prediction of
Unsteady Aerodynamic Flows
Dr. Siavash Kassemi
44
Investigation of Lithium in Buffered
MEIC-AlClj Melts
Dr. Bernard Piersma
45
Control Formulations for the Active and
Passive Damping of Flexible Strucmres
Dr. Thomas Posbergh
Geophysics Laboratory
46
Background Research on Global Warming
Dr. Theodore Aufdemberge
47
Correlations Between Levels for Stellar
Scintillometer Derived Profiles of C^
Dr. Frank Battles
48
Total Dose Effect on the Soft Error Rate
of Metal-Oxide-Semiconductor Memory Cells
Dr. Reuben Benuroof
xUv
Volume n
49
PEL Short-Wave Disturbances over the
Desert Southwest
Dr. Chia-Bo Chang
50
Simulation of REFS Missile Flight
Dr. George Jumper
51
Evaluating the Diagnostic Potential of
High Spatial and Spectral Near Infrared
Observations of the Solar Photosphere
Dr. Jeffrey Kuhn
52
A New Ion-Molecule Chemiluminescence
Experiment
Dr. C. Randal Lishawa
53
Reladonship Between Brightness
Temperatures and Typhoon Intensification
Dr. Gandikota Rao
54
Electric Fields in the Middle-and-Low
Latitude Ionosphere and Plasmasphere
Dr. Craig Rasmussen
55
Resonance Enhanced Multiphoton Ionization
of Molecular Nitrogen/Electronic Quenching
of the Nj A State by CO
Dr. Glenn Stark
56
Optical Propagation in Non-Uniform Media
Dr. John Wills
Rome Air Development Center
57
Implementation of ACT Adaptive Filters
Dr. Charles Alajajian
58
Exploiting Parallel Architectures within
a Distributed Computational Environment
Dr. Gary Craig
59
Optical Simulations of Guided-Wave
Structures
Dr. Lionel Friedman
60
Magnetically Controllable Microstrip
Path Analysis
Dr. Frances Haiackiewicz
61
Scattering by Conductor-Backed Dielectric
Gaps
Dr.Hao Ling
62
An Efficient Parallel Algorithm and Its
Implementation for Real-Time Adaptive
Space-Time Processing
Dr. Shietung Peng
xlv
Volume II
63
Processing and Characterization of Pb-doped
Bi-Sr-Ca-Cu-0 Superconducting Thin Films
by the MOD Method
Dr. Daniel Ryder
64
Architectural Support for AI and Knowledge
Base Systems
Dr. Behrooz Shirazi
65
Markov Models for Simulating Error Patterns
on Data Communications Links
Dr. Wayne Smith
66
Use of Audio Feedback to Confirm Verbal
Commands for Computer Workstations
Mr. Ronald VanEtten
67
Theoretical Models of Fast Photoconducting
Avalanche Switches
Dr. Martin Wilner
68
A Gabor Transform Based Recognition System
Dr. James Wolper
Weapons Laboratory
69
Palindrome Pre-Scheduling
Dr. William Campbell
70
Second-Harmonic Generation in Corona-Poled
Polymer Films
Dr. Gene CarUsle
71
Application of the Microplane Concrete
Model to an Explicit Dynamic Finite
Element Program
Dr. William Cofer
72
From Counterpropagation to Vector
Quantization: Neural Networks for Pattern
Recognition
Dr. Johanna Schruben
73
Analysis of Data on Compact Toroid
Formation in Hydrogen
Dr. Peter Walsh
74
AOA Determination using Associative Neural
Networks
Dr. Wayne Zimmerman
xlvi
Volume in
(Wright Research Development Center)
Aero Propulsion Laboratory
75 Evaluation of MOS-ConuoUed Thyristor (MCT)
at 270 Volt DC for Resistive and Inductive
Loads
76 Development of a Three-Dimensional Finite-
Difference Code for Modeling Flow and
Heat Transfer in Rotating Disk Systems
77 Thin Film Behavior of Powder Lubricants
Mixed with Ethylene Glycol
78 Laser Velocimetry Measurements in Shock
Tubes
79 Thermal Analysis of Potential Solid
Lubricant Candidates
80 Effect of Evaporation on the Driving
Capillaiy Pressure in Capillary Pumped
Aerospace Thermal Management Systems
81 Investigation of the Combustion
Characteristics of a Confined Coannular
Jet with a Sudden Expansion
82 Aircraft HVDC Power System - Stability
Analysis
83 Design of a Dynamic Temperature Measurement
System for Reacting Rows
84 Hydrogen Permeation in Metals at Low
Temperatures
85 Measurements of Droplet Velocity and Size
Distributions for a Pressure/Air Blast
Atomizer
Avionics Laboratory
86 Pattern Recognition: Machine vs. Man
Dr. Muhammad Choudhry
Dr. Mingking Chyu
Dr. Donald Dareing
Dr. Paul Dellenback
Dr. Dennis Flentge
Dr. Kevin Hallinan
Dr. Paul Hedman
Dr. K. Sankara Rao
Dr. Larry Roe
Dr. Kaveh Tagavi
Dr. Richard Tankin
Mr. Thomas Abraham
xlvii
Volume in
87
Some Results in Pattern-Based Machine
Learning
Dr. Michael Breen
88
Probabilistic IR Evidence Accumulation
Dr. R. H. Cofer
89
Investigations of a Lower Bound on the
Error in Learned Functions
Dr. Thomas Gearhart
90
Machine Learning Applied to High Range
Resolution Radar Returns
Dr. Lawrence Hall
91
Model for Characterizing a Directional
Coupler Based Optical Heterodyne
Detection System
Dr. Mohammad Karim
92
Context Dynamics in Neural Sequential
Learning
Dr. Kevin Kirby
93
Fiber Laser Preamplifier for Laser Radar
Detectors
Dr. Richard Miers
94
Reusable Ada Software • Evaluating the
Common Ada Missile Packages (CAMP-3)
Dr. Brian Shelburne
Electronic Technology Laboratory
95
Computer Simulation of NMOS Integrated
Circuit Chip Performance Indicators
Dr. Ashok Goel
96
Application of Photoreflectance to Novel
Materials
Dr. Muhammad Numan
97
Electronic Structure and Deep Impurity Levels
in GaAs Related Compound Semiconductors and
Superlattices
Dr. Devki Talwar
Flight Dynamics Laboratory
98
Sensor Integration Issues in Robotic Rapid
Aircraft Turnaround
Dr. John Bay
99
Influence of Static and Dynamic Aeroelastic
Constraints on the Optimal Structural
Design of Flight Vehicles
Dr. Franklin Eastep
xlviii
Volume in
100
Location of Crack Tips by Acoustic Emission
for Application to Smart Structures
101
H„ Design Based on Loop Transfer
Recovery and Loop Shaping
102
A Feasibility Study on Interfacing Astros
with Navgraph
103
Theoretical Modeling of the Perforation of
Laminated Plates by Rigid Projectiles
104
Accelerate Fatigue Test Procedure for the
Structural Polycarbonate Component of the
F-16 Canopy Composite Material
105
Study of Fracture Behavior of Cord-Rubber
Composites ^or Lab Prediction of Structural
Durability c . Aircraft Tires
106
Ballistic Damage of Aircraft Structures:
Detection of Damage Using Vibration Analysis
*** Submitted as Technical Memorandum
107
Delamination of Laminated Composites
108
Experimental Identification of Internally
Resonant Nonlinear Systems Possessing
Quadratic Nonlinearity
Materials Laboratory
109
The In-situ Laser Deposition of High T^
Superconducting Thin Film
110
AMI Calculations on Rigid Rod Polymer
Model Compounds
111
Potentials of Mushy-State Forming of
Composite Materials
112
Structural Analysis of Polymer Precursors
with Potential Nonlinear Optical Properties
Dr. Marvin Hamstad
Dr. Chin Hsu
Dr. Ming-Shu Hsu
Dr. David Hui
Dr. Yulian Kin
Dr. Byung-Lip Lee
Dr. Vernon Matzen
Dr. William Wolfe
Dr. Lawrence Zavodney
Dr. Donald Chung
Dr. John Connolly
Dr. Sherif ElWakil
Dr. David Grossie
xlix
Volume in
113
Eddy Current Testing in Nondestructive
Evaluation
Dr. Prasad Kadaba
114
Preparation and Characterization of
Polypeptide Thin Films
Dr. Joseph Lambert
115
Chemical Induced Grain Boundary Migration
in AI2O3
Dr. Gary Leatherman
116
On the Use of QPA (Qualitative Process
Automation) for Batch Reactor Control
Dr. Won-Kyoo Lee
117
Ultrasonic Techniques for Automated
Detection of Fatigue Microcrack
Initiation and Opening Behavior
Dr. Michael Resch
118
NMR and IR Investigations of Conformational
Dynamics and Surface Interactions of
Pcrfluoropolyalkylethers
Dr. Martin Schwartz
119
Modeling of Casting Solidification
Dr. Hai-Lung Tsai
1
Volume IV
Human Systems Division Laboratories
Harry G. Armstrong Aerospace Medical Research Laboratory
120 Cardio-Respiratoiy Measures of Workload
During Continuous Manual Performance
121 Heat Transfer Through Multiple Layers
of Fabric
122 Pilot Task Functional Analysis and
Decomposition Using Structured Analysis
and IDEF Modeling Methods for the Pilot’s
Associate Pilot- Vehicle Interface
123 Effects of Time Delays in Networked
Simulators
124 Speaker Normalization and Vowel Recognition
using Neural Networks
125 Sensitivity Analysis of the PB-PK Model:
Methylene Chloride
126 Enror Analysis of the AAMRL Inertia
Testing System
127 Simulation of Head/Neck Response to -Gx
Impact Acceleradon
128 Improving Pilot Efficiency in the Age of
the Glass Cockpit: Designing Intelligent
Software Interfaces for the Military
Aviation Setting
129 Decision-Making Under System Failure
Conditions
Human Resources Laboratory
130 An Intelligent Tutoring System to Facilitate
Invention Strategies for Basic Writing
Students
131 A Comparative Analysis of a 4-Group and
6-Group Job Classification
li
Dr. Richard Backs
Dr. Larry Byrd
Mr. John Duncan
Dr. Martin Hagan
Dr. Ashok Ktishnamimhy
Dr. Tze San Lee
Dr. Sigmund Lephart
Dr. Amit Patra
Dr. Leonard Shyles
Dr. Bonnie Walker
Dr. Margaret Batschelet
Dr. Pinyuen Chen
Volume IV
Dr. James Dykes
132 Optimizing the Training and Acquisition
of Complex Spatial Skills
133 Decision Processing in Dynamic Decision
Environments
134 The Use of CAD to Develop ICAI for the
Improvement of Spatial Visualization Skills
135 Considerations in the Assessment and
Evaluation of Mental Models
(Technical Memorandum)
136 Automating the Administration of USAF
Occupational Surveys
137 Psychophysical Measurement of Spectral
Attenuation in the Human In Vivo Ocular
Media: Method and Results
138 Benefit-Cost Evaluation of Simulator Based
Multiship Training Alternatives
139 Determinants of Staying and Leaving of
Military Medical Personnel From a US
Air Force Hospital
140 Cognitive Representations of Teams
141 Recruit of the Year 2(XX) and the Fundamental
Skills
142 Survival Analysis: A Training Decision
Application
143 Predicting the Impact of Automation on
Performance and Workload in C? Systems
Occupational and Environmental Health Laboratory
144 An Assay to Determine the Phytotoxic Effects
of Jet Fuel: Effects on Vesicular-
Arbuscular Mycorrhizae
Dr. Daniel Garland
Mr. Harold Goldstein
Dr. Verlin Hinsz
Dr. Delayne Hudspeth
Dr. Gillray Kandel
Dr. William Moor
Dr. James Price
Dr. Joan Rentsch
Dr. Eleanor Smith
Dr. Stanley Stephenson
Dr. Pamela Tsang
Dr. David Buckalew
lu
Volume IV
145
Mathematical Modeling and Decision-Making
for Air Force Contaminant Migration Problems
Dr. Miguel Medina
146
An Assessment of Hazardous Waste
Minimization Efforts in the United States
■ Air Force
Mr. Kirk Nordyke
147
Beam Profile Characteristics of the Shephard
Cs-137 Gamma Irradiator at the AF
Occupational & Environmental Health Laboratory
Instrumentation Calibration Facility Brooks AFB
Ms. Lorin Weber
School of Aerospace Medicine
248
Comparisons of Air and Liquid
Microenvironmental Cooling
Dr. Phillip Bishop
149
Image Analysis of Raw Macrophage Cells
Dr. Robert Blystone
150
Perception and Attention in Three-Dimensional
Visual Space
Dr. Bruno Breitmeycr
151
PCR Analysis of Ureaplasma urealyticum
and Mycoplasma hominis
Dr. Vito DelVecchio
152
The Effect of Absolute Humidity on
Thermoregulation by Rhesus Monkeys
Dr. Randall Dupre
153
Effects of Microwave Radiation on Yeast
Cells
Dr. Reinhard Graetzer
154
Determination and Analysis of Range Data
Using Computer Vision
Dr. Paul Griffm
155
Dioxin Half-Life Estimation in Veterans of
Project Ranch Hand
Dr. Pushpa Gupta
156
A Comparison of Various Estimators of
Relative Risk in Epidemiological Studies
Dr. Raraesh Gupta
157
Predisposition of Mammalian Cell Cultures
Treated with Aflatoxin B1 to Potential
Radiation Effects
Dr. Paul Lemke
158
No Report Submitted
Dr. Rex Moyer
liii
Volume IV
159 The Effect of Hyperbaric Oxygenation on
Denervation Induced Muscle Atrophy
160 Bioeffects of Microwave Radiation on
Amino Acid Metabolism by RAW 264.7
Mouse Macrophage Cells
161 Neural Graft-Host Brain Interactions
Visualized with Voltage-Sensitive Dyes
162 Development of an Enhanced Hydraulic
Cardiovascular Model/Test Apparatus for
In-Vitro Simulations in Altered-g
Environments
163 Pulmonary Measurements in Hyperbaric and
Non-Hyperbaric Exposures Ad^ndum to:
The Reduction of Denervated Atrophy as a
Consequence of Hyperbaric Oxygen Treatment
164 Characterization of +Gz- Induced Loss of
Consciousness in Rats
Wilford Hall Medical Center
165 Interrelationships of Tobacco, Caffeine,
and Alcohol Use Among Participants of an
Air Force-Sponsored Health Promotion Program
Mr. Arnold Nelson
Dr. Donald Robinson
Dr. David Senseman
Dr. Richard Swope
Dr. John Szarek
Dr. Steven Waller
Dr. Janet Dizinno
liv
1990 USAF-UES SIM4ER EACULTy BESEARCH PROGRAM/
GRADUAIIE STODEtTF RESEARCH PROGE^
{Sponsored by Ihe
AIR PCRCE CaSFICE OF SCIENTIFIC RESEARCH
Conducted by the
Universal Energy Systems, Inc
FINAL REPORT
Evaluation of MOS-Controlled Thyristor (ICT) at 270 Volt DC
for Resistive and Inductive Loads
Prepared by:
Academic Rank:
Department and
University:
Mjhamn»d A. Choudhry, Ph.D,
Professor
Department of Electrical and Ccnputer Engineering
West Virginia University
Research Location: Electricjil Technology/Aerospaoe Power Division
Wright”Patterson AFB
Ohio 45433-6563
USAF Researcher:
Date:
Contract No:
Joseph A. Weimer
August 15, 1990
F49620-88-C-0053
Bvzduation of MDS-Corxtrolled Thyristor (ICT) at 270 Volt DC
for Resistive and Inductive Loads
hy
Muhanmad A. Choudhry
ABSTRaCT
Ohe voltage and current characteristics of MDS-Controlled Thyristors of
50 to 150 A rating are obtained at 270 volt dc over a wide range of
switching frequencies. Total turn-off time of MCT increases with load
current and switching frequency aixl is less than 2 u sec in most cases. The
forward voltage drop across M3T varies frcm 1 volt to 2 volts for a load
current variation of 15 to 90 A. Large voltage transients are observed
during switching of inductive loads. The use of snubber circuit across ICT
reduces voltage transients and power loss inside the device. However, large
currents are observed during tum-on of MCT at hi^ frequency with inductive
load.
75-2
AdoKfwledgments
I wish to thank the Air Force Systems Cotinand and the Air Force office
of Scientific Research for sponsorship of this research. Ihe help provided
by Greg Fronista, Joseph P. Walick Jr., Michelle Mneller, Bick Nguyen, John
Nairtjs, Samuel W. Sexton, Philip C. Herzen, Lawrence C. Walko, Vic MctTier
and Gary Nddb was invaluable in ccnpleting this research. Joseph A. Weimer
provided an enjoyable worJdng enviromient. Ihe discussions with Lt. Ihcmas
D. King, Miguel A. Maldonado, Clarice W. Severt and Rene J. Thibodeaux were
very helpful. The interest of L. Dave Massie, Tern Mahefky, Maj Michael D.
Braydich and William U. Bozger in research work was greatly appreciated.
The help of Ikiiversal Energy System in administering this contract is also
acknowledged.
I. INTBODOCTICN:
Uhlike Gate-Tum- Off Thyristor which can be turned on ty applying a
small positive pulse to the gate but requires about one fifth of normal
current to turn off, MDS-Controlled Thyristor can be turned on and off with
a very small amount of current pulse applied between gate and anode. The
Electrical Technology Section of Aero.«;paoe Power Division at Wiright-
Patterson AEB is interested in application of MCT in "More-Electric
Aircraft" due to its light“w®igi*/ low forward voltage drop and large
current carrying capability.
tty research interests have been in the areas of control and analysis of
ac/dc power system, offset of nonlinear load on ac/dc power system stability
and use of power electronics to enhance safety and increase productivity in
underground coal mines, tty work on the use of GTO to svppress arc in mine
dc haulage systems and analysis of ac/dc power system contributed to ity
assignment to Electrical Technology/Aerospace Power Division.
75-4
II. CBIECTIVES OF THE PESERBCH EFEX»T:
Power electronic devices can be used for control and efficient
xitilization of electrical energy. However, these devices may be damaged
the electrical transients due to switching of the device or the load. MDS-
Controlled Thyristor (ICT) can be turned on and off with a very small amount
of current pulse aqpplied between anode and gate. This device has low
forward voltage drop, hi^ current czpability and low weig^it/power ratio.
My assignment as a participant in the 1990 Sumnter Faculty Research
Program (SERP) was to evaluate voltage and current characteristics of tCT at
different switching frequencies with different types of loads at 270 V DC.
This information will be utilized in the "More-Electric Aircraft" initiative
>hich will inprove thrust to wei^it ratio, reliability and sedfety of the
aircraft.
III.
a. MDS-Controlled Thyristor has hi^ current density, high dl/dt capability
and very small tum-on and turn-off time [!}. This device is expected to
displace SCRs and GTO in the future. Future aircrafts are e:pected to use
more electric power at 270 V DC to inprove reliability and safety. Figure 1
shows the single line diagram of the circuit used for evaLLuating the
voltage/current characteristics of MCT at 270 VDC. Three-phase ac output of
auto-transformer is rectified and a capacitor bank is connected at the
rectifier to reduce the ripple in the dc voltage. A large resistor is
connected parallel to the capacitor bank to discharge the c^jacitors vhen
75-5
Thrae-phase
supply
440 V
1-1
switcii SW is opened. A filtering capacitor is connected at the ICT to
reduce the voltage variation of the anode voltage during tum-on and turn¬
off of ICT. A free-vdieeling diod is connected across the load to reduce
switching transiait voltage across ICT. Resistive load was modoled by glow
bar and water-cooled resistors.
b. Figure 2 shows the variation of on-state forward voltage drop in ICT
with anode current. Ihe on-state forward voltage drop decreases with higher
junction tenperature of ICT. Figure 3 shows the voltage waveform across ICT
at a switching frequency of 10 K Hz for a resitive load at 270 VDC. Figure
4 shows the e:panided voltage waveform during tum-on and turn-off of ICT.
The voltage across ICT decreases from 90% to 10% of dc input voltage in 222.3
u sec.
Figure 5 shows the anode current and gate current waveform at a
switching frequency of 10 K Hz for a resistive IozkI at 270 VDC. A gate
current of less than 2A is required to txim-on and turn-off 38. 2A throu^
the anode of ICT. Figure 6 shows the expanded anode current waveform during
tum-on and turn-off of ICT. The current increases eaponentially during
t\im-on due to a small inductance present in the load circuit. The total
turn-off time (decreasing of anode current to 10% of its value before
turn-off after the gate voltage rises through zero) is 1.92 u sec. Table 1
gives the toted, turn-off time of rCT as a function of anode current and
switching frequency. The turn-off tiine increases with load current and
switching frequency.
75-7
0
>
•0
(0
X
u
0
bi
1.0
0.5
0
Figure
(100 volta/div)
-100.000 US 0.00000 S 100.000 US
Ficruro 3. Voltage waveform across MCT at a switching frequency of
10 KHz for a resistive load at 270 v DC.
75-9
ClOO volts/ div) V .(100 volts/ div)
2.50000 US
0.00000 S
a.soooo US
d
>
Figure 4. Voltage waveform during turn-on and turn-off of MCT for
a resistive load at 270 v DC.
75-10
Gate current (2 A/ div) Anode current (20 A/ div)
75-11
Gate current (2 A/ div) Anode current <20 A/ div)
Figure 6. Anode current waveforms during turn-on and turn-off at
switching frequency of 10 KHz for a resistive load at
270 V DC.
Copy io DTiC does no!
pprmif * '1/ >■, ri Dfoduption
75-12
Table 1 . Total turn-off time of MCT as a function of anode current
and switching frequency.
Switching Frequency* 400 Hz
Switching Frequency* lOKHz
Anode Current
Total turn-off time
Anode Current
Total turn¬
off time
(A)
( sec)
(A)
( sec)
10.97
1.32
10.74
1.36
20.34
1.39
17.83
1.51
36.47
1.44
42.86
1.98
75-13
Figure 7 shows the voltage, gate and anode current waveforms of MCT for
an inductive load at a switching frequency of 10 K Hz at 270 VDC. A large
transient current flows through MJT during tum-on vdrLch causes voltage
oscillations across tCT. After the initial current, the current increases
exponentially at a time constant determined by the inductance and resistance
of the load circuit.
Figure 8 shows the voltage waveform across HJI without the filtering
capacitor at the HCT at a switching frequency of 400 Hz. The input voltage
is 100 VDC. A transient voltage of 263.6 V appears across M3T. Figure 9
shows the voltage waveform across MDT with a snubber circuit at a switching
frequency of 400 Hz. The input voltage is 270 VDC. A transient voltage of
381.8 V spears across the device. The use of smhber circuit reduces the
voltage transient across £CT and reduces the switching power loss inside
MCT.
IV. BBOCWMEIEIATICNS;
The MOS-Controlled Thyristor is able to turn-off large currents in 3.ess
than 2 u sec at 270 VDC. However, large voltage oscillations spears cJuring
tum-on with inductive load at higher switching f requencrles . This research
shows that the nature of the load has significant effect on the performance
of MCT. Further research will be done to evaluate voltage/current
characteristics of MCT in series with a dynamic load. Ccnputer simulation
will be develcped to optimize the design of snubber circuit of tCT and to
determine the cause of high surge current during t\im-on at hic^ switching
frequency.
75-14
Gate current Anode current V . (100 volts/ div)
(2 A/div> <10 A/div) _
AU1i 0.000U
TRIG 2 = 0. 13U
yAT= V
0.1U
IQ.
U
PERKDE
K 20
ix S
Firgure 7 .
/oltage and current waveforms for an inductive load at
a switching frequency of 10 KHz at 270 v DC.
(100 volts/ div)
Figure 8 . Voltage across MCT without a filtering capacitor at MCT
for an inductive load at 100 v DC.
75-16
(100 volts/ div>
75-17
1. V.A.K. Tenple. lOa - Thyristors for the Future. Power technics
Magazine, NOventoer 1989, pp. 21-24.
75-18
1990 USAF-UES SUMMER FACULTY RESEARCH PROGRAM/GRADUATE
STUDENT RESEARCH PROGRAM
Sponsored by the
AIR FORCE OF SCIENTIFIC RESEARCH
Conducted by the
UNIVERSAL ENERGY SYSTEMS, INC.
HML.REPQRT
DEVELOPMENT OF A THREE-DIMENSIONAL FINITE-DIFFERENCE CODE
FOR MODELING ROW AND HEAT TRANSFER IN ROTATING DISK SYSTEMS
Prepared by:
Academic Rank:
Mingking K. Chyu
Associate Professor
Department of Mechanical Engineering
Carnegie Mellon University
Pittsburgh, PA 15213
Research Location: Wright Research and Development Center,
POOS-3, Dayton Ohio
USAF Researcher: Dr. Won S. Chang
Date:
Contract Number:
September 30, 1990
F49620-88-C-0053
DEVELOPMENT OF A THREE-DIMENSIONAL FINITE-DIFFERENCE CODE
FOR MQDElTNq-£LQ.W.ANp ?iEAI.TRANSFgHIN.RQTATING.PISK SYSTCMg
M. K. Chyu W. S. Chang
Dept, of Mechanical Engineering AFWAL/POOS-3
Carnegie Mellon University Wright Research Development Center
Pittsburgh, PA 15213 WPAFB, OH 45433
ABSTEACI
A pressure-based, three-dimensional, finite-difference (finite-volume) code for
modeling the fluid flow and heat transfer in a ( tine disk system has been developed. The
computer program developed, however, is capuole of solving the governing equations for
convection transport in elliptic nature. The computational procedure employs a colocated
grid system extended for generalized non-orthogonal coordinates thus can handle irregular
geometries. The velocity-pressure coupling uses a modified version of the pressure implicit
split operator (PISO) originally proposed by Issa in 1985. The PISO algorithm not only
gives time-accuracy results but also provides the computation with a much faster
converging speed as compared to the SIMPLE based counterparts. Ongoing research using
this computational program is to investigate the effects of radial and axial clearance on the
heat transfer in a turbine rotor-stator system. Near-future study plan will focus on counter¬
rotating disk systems which appear to be the viable configurations used for future turbine
engines.
76-2
ACKNOWLEDGEMENTS
The first author (MKC) of this report is grateful to Air Force Systems Command, Air Force
Office of Scientific Research, Universal Energy System and AFWAL/POOS-3 for the
appointment of Summer Faculty Research Fellowship. In addition, he expresses his special
appreciation to Dr. Won S. Chang at AFWAL/POTC for his collaboration in this project .
76-3
NOMENCLATURE
m
geometric coefficients
j
Jacobian
p
pressure
Pr
Prandtl number
S
source term
T
temperature
ui
velocity at duct inlet or velocity components (u,v,w)]
in cartesian coordinate (x,y,z)
Ui
contravariant variables (U,V,W)
x,y,z
cartesian coordinates
X,Y,Z
(x,y,z) normalized by D, see Fig. 2
viscosity
P
density
general non-orthogonal coordinate system
Subscript
P.E.W.WW
variable values at node P, E...at the discretized domain
Superscript
n
at current time step
n+1
at next time step
*
i
at intermediate time step
764
1. INTRODUCTION:
The study of the heat transfer characteristics of rotating disk systems is a subject of
importance in the design of various types of rotating equipment, particularly for the
aeropropulsion and turbomachinery. In the continual quest for improved performance and
efficiency, modern gas turbine designers strive to develop systems and/or materials that
will permit increasing turbine operating ternperamres. Consequently, much effort has been
expended to enhance understanding of the hydrodynamic and heat transfer phenomena of
various types of rotating disk systems.
Heat transfer from a disk rotating in a quiescent environment has been examined in
numerous experimental and theoretical studies. Results of such investigations (Millsaps
and Pohlhausen (1952), Cobb and Saunders (1956), Sparrow and Gregg (1960), and Ong
(1988)) have provided extensive insight into the hydrodynamic and thermal characteristics
of free disks. While these results may be of limited applicability in turbine design, they
often serve as a baseline for gauging the effectiveness of various forced cooling techniques
and forjudging the heat transfer characteristics of rotor-stator systems, which are more
appropriate idealized turbine disk models.
Unlike the free disk in which heat transfer is only affectea by rotational speed, heat
transfer in rotor-stator systems is influenced by the operating conditions as well as system
geometry. In a rotor-stator system with no radial shroud, commonly referred to as an open
rotor-stator, Kreith et. al. (1959, 1963) and Metzger (1970) demonstrated that the presence
of the stator influences heat transfer from the disk at axial gap spacings less than 10 percent
of the disk diameter. These studies also examined the effects that forced coolant flow
introduced at the hub of the system had on the average heat transfer rate. As an extension
of these studies, more recent evaluations by Metzger et al. (1979, 1989) and Popiel and
Boguslawski (1986a& 1986b) have examined the effects of varying the location and rate of
forced coolant flow. In these studies the effects coolant jet impingement have on rotor local
heat transfer coefficients were invesrigated.
Another important geometrical effect in actual turbines is the presence of a radial
clearance seal. While the primary intent of forced disk cooling systems is to enhance heat
transfer from the rotating turbine disk, this flow is also intended to prevent the ingress of
hot combustion gases. Phadke and Owen (1983) investigated the hydrodynamic effects of
different seal designs to assess their ability to prevent ingress, but few studies have
76-5
explicitly examined the effect that a radial clearance shroud has on heat transferrates.
Metzger et. al. (1989) performed local heat transfer measurements across a disk face using
a radially shrouded rotor-stator system, but the purpose of these studies was to investigate
the influence that the radial location and rate of coolant flow have on the local heat transfer
rates.
II. OBJECTIVES OF THE RESEARCH EFFORT;
The primary objecrive for this summer research is to establish a numerical code for
modeling convection transport phenomena in turbine disk systems. The computational
procedure developed should, at least, have the capability of addressing the important issues
and research concerns involved in the rotating disk problems. Moreover, it is targeted to
become one of the most powerful and modem computer programs for general elliptic
problems in the area of computational fluid dynamics (CFD) and heat transfer. Many
special features and algorithms of recent advent in the CPD community are also included in
the procedure. With an extensive review and numerical testing, the final ramification is a
pressure-based, finite-difference code capable of solving the incompressible, three-
dimensional, full Navier-Stokes equations in non-orthogonal coordinates. One special
feature of this finite-difference (finite-volume) procedure is the capability of dealing with
irregular geometries, which is used to be monopolized by the finite-element methods. In
addition, it adopts a non-staggered grid arrangement, thus the velocity components and
other dependent variables are colocated at the same grid. The velocity-pressiu^e coupling is
resolved by the pressure implicit split operator (PISO) (Issa, 1985) with an extension for
generalized non-orthogonal coordinates. Discretization of the governing equations uses the
central-differencing scheme for the diffusion terms and the second-order upwind scheme
for the convection terms. The solution of the discretized equations is obtained from a
conjugate gradient method (Kershaw, 1978), instead of the conventional tri-diagonal matrix
algorithm (TDMA) incorporated with a line-by-line iteration. The remaining part of this
report describes the framework of the code.
III. MATHEMATICAL FORMUALTION:
For laminar flow of Newtonian fluid with constant properties, the governing
equations for the unsteady flow and heat transfer in cartesian tensor form are
Continuity
76-6
(1)
9ui
= 0
9xi
Momentum
9pui 9 , .
+ — (pUiUj)
9t 9xj
Energy
9pT 9 /
— (pujT)
9t 9xi
A + i_
9xi 9xj
p
^9xj 9xi/
9xj
r axj
(2)
(3)
To treat arbitrary boundaries, these governing equations are transformed to a general form
based on a non-orthogonal coordinate system (^, q, Q. There art three different ways to
represent velocity field in a non-orthogonal grid system; i.e., (1) cartesian velocity
components (Rhie and Chou, 1983; Han, 1988), (2) contravariant velocity components
(Maliska and Raithby, 1984) and (3) resolute or velocity projections along the coordinate
directions (Karki and Patankar, 1989). The present transformation uses the cartesian
velocity components, which has the advantage of excluding the so-called "curvature" terms
(Rhie and Chou, 1983). Using the standard transformation formulae (Anderson et al.,
1984), the continuity equation becomes
9U 9V 9W
9^ ^ 9n 9C
(4)
and the remaining governing equations for the dependent variable (}) in the general
coordinate system can be expressed in a compact form as follows;
+ ^(pU(}>) + |;-(pV<t>) -i- ^(pW(}>)
d_
ro
^(gll + g^2 + gis)
an
y(g2i + g22 + glsl^j
J 9n.
(5)
76-7
where U, V, and W are the contravariant variables that represent convective flux ;
U = giiu + gi2V + gi3Wj
V = g2lU + g22V + g23W orUi = gijUj (6)
w = g3lU + g32V + g33w)
and r(() and S'!* are the associated diffusivity and source terms for the variable (])
(=u,v,w,T). The detailed expressions of the source terms and the cross derivative terms
due to grid non-orthogonality have been given by Han (1988). The geometric coefficient
gij and the Jacobian J are defined as
X^Z^ - Xt^Z^
xtiy^-x^y^ ■
gij =
y^z^ - y^z^
x^z^ - x^z^
x^y^ - x^y;
Hyr\ - XTiy^
J = + X^^Z;^
+ Xi^y^^ -
- W?
and
(7)
These geometric coefficients are evaluated using central differencing schemes in the
transformed domain.
IV. DISCRETIZATION:
Discretization of the governing, differential equations uses the finite-volume
approach. Differencing in the temporal domain employs the Implicit Euler scheme. All the
dependent and independent variables are stored at the same grid location and variables at the
finite control-volume boundaries are interpolated between adjacent grid points. In the
spatial domain, the diffusion terms and convective terms are approximated by the central
differencing and a second-order upwind scheme, respectively. The present upwind scheme
determines the prevailing value of a dependent variable at the finite-volume interfaces by
referring to the sign of the contravariant fluxes. For example, if the contravariant flux U at
point p is positive, the finite- volume face value (j)w (west to P) is calculated by linear
extrapolation from the corresponding variable at the two closest upstream neighboring
points (ww means further west to the west). This leads the discretization of the term
— Ap(t)p - - A\vw<l>vw
where
A _ Axv^ (2Axww - Ax^) + Axg (Ax^w - Axv^) Ue
2Ax\^ (Ax\^\y * Ax^)
Aww ~
- Axw'Uw
2 (Axww ' Axw)
and
An — Aw + A
ww
V. VELOCITY-PRESSURE COUPLING:
A typical computational procedure for compressible flow employs density as a
primary variable and extracts pressure from an equation of state. This density-based
method is difficult for incompressible or low-Mach number flows, since weak pressure-
density coupling prevails under these situations. Computation for incompressible flow is
generally performed using pressure-based method which is characterized by the use of
pressure as one of the primary dependent variables. However, the absence of both
unsteady term in the continuity equation and an explicit equation specifically for pressure
imposes that the pressure can only influence the velocity field through the momentum
conservation with the continuity as a compatibility condition. Hence provision of an
algorithm for the velocity-pressure coupling is essential. One well-known algorithm
employs semi-implicit iterative coupling procedure; e.g., SIMPLE (Patankar, 1981) and its
variants (Van Doormaal and Raithby, 1985). In the present study, a more efficient non¬
iterative method based on the operator-splitting technique (PISO) (Issa, 1985) is extended
to non-orthogonal grid coordinates. The discretized governing equations are solved in a
time-marching fashion until a steady state solution is reached. In each time step, a two-
stage correction procedure follows an initial predictor stage, and all three stages combined
yields a time-accuracy result. The sequence of the scheme is as follows:
a) Predictor Step:
76-9
Using the pressure field at time step t’', the discretized equation for ui is
^ { (pui)* - (pui)" ) = H(ui ) - gji Vj p" + Si
where H( ) represents the finite difference discretization operator. With
h{u*) = ntu*) + Ap u* ^ the first step becomes
-2- - Ap u- = H'(u*) - gji Aj pn + Si + -^uf (9)
\At / At
The coefficients Ap's involve contributions due to the convective fluxes as well as diffusive
fluxes in the T), ^ directions.
b) First Corrector Step
The mass fluxes calculated from equation (9) unnecessarily satisfy continuity, the
first corrector step is established by splitting the operator
H(up) = ApUp + Htu*), thus
^ - Ap)ur = H'(ut) - gjiVjp* + Si + ^u?
(10)
Subtracting equation (9) from (10) and imposing the continuity equation (Vj Uj** = 0), the
pressure correction equation is resulted by taking divergence of the incremental momentum
correction equation:
Ap g/i Vj (P* - P")
- Vj gji Ui - Vj Uj
(11)
Note that the first corrector step is similar to the SIMPLE algorithm. In fact, if the loops
are iteratively terminated here, it is identical to the SIMPLE method with the time-step-
involving term as an under-relaxation factor. To account for the effects of convection and
diffusion and to satisfy the momentum conservation, a second corrector step follows.
76-10
Further splitting the operator; i.e.,
hU;*') = Ap + H'lur),
the momentum equation becomes
Subtracting equation (10) from (12) gives the incremental momentum equation. Then, by
invoking the continuity again, the pressure increment equation becomes:
The splitting errors after two corrector stages are less than the truncation error associated
with the Euler implicit approximation in the temporal domain. The prevailing values
obtained here are taken to be the values at the next time-step,
It is known that the non-staggered grid system can produce the undesirable checker¬
board phenomena in the pressure field, due to the decoupling effect between velocities and
pressures. As suggested by Rhie and Chou (1983), an explicit fourth order pressure
damping terms may be added to the right hand side of equation (1 1) to ensure an attainment
of oscillatory-free pressure field. This approach can be further modified to include the
transient term for time-marching formulations. However, the steady-state solution obtained
in this way is found to be strongly influenced by the size of time step chosen, which is
undesirable (Parameswaran and Sun, 1989). In this study, the inclusion of the transient
term in the formulation is done by the operator interpolation technique first proposed by
Peric (1988). For example, the Uj's evaluated at the cell interface i+ 1/2 becomes:
" f [l^ ■ ' ("bi-i) + s,., + iu?.,)
76-11
+ gji (P?+l - I?) (14)
The resulting mass fluxes are then substituted into the right hand side of pressure correction
equation (11) and (13).
The system of finite difference approximation equations formed by equations (8),
(1 1) and (13) produces a seven-banded matrix. For both pressure correction equations, the
matrix is symmetric and positive-definite. To enhance the matrix solver efficiency, the
conjugate gradient (CG) method of Kershaw (1978) is employed for reducing the residuals
of these two equations to 10'3.
In actual computarion, it generally solves the non-dimensionalized governing
equations. Dependent variables are scaled to the characteristic lengths, velocities and
corresponding references. A preliminary result for a study concerning the radial and axial
clearance effects on the heat transfer in a rotor-stator system shows that the present code
performs far superior than the SIMPLE related counterparts, by a factor of 10 - 100 in
converging speed. Most of the pressure-based codes in the public domain or commercially
available are based on SIMPLE algorithm or its derivatives.
VI. RECOMMENDATION:
Besides the ongoing research for the rotor-stator system, this code is capable of
performing similar modeling for other disk configurations. Because its feature of
formulating governing equation in generalized non-orthogonal coordinates, the effects of
geometric details; e.g. uneven thickness and surface curvature, can be studied with a grid
generation. These issues have never been explored in the past. Another important
configuration is the system involving counter-rotating disks, that is possibly to be used in
the future high-performance turbine engines. Many existing theories and findings for the
conventional rotor-stator system are inapplicable for the counter-rotating disks; both
numerical modeling and experimental work for code validation are strongly desirable.
76-12
REFERENCES
Anderson, D.A., Tannehill, J.C. and Fletcher, R.H., Computational Fluid Mechanics
and Heat Transfer, Hemisphere Publication, Washington D,C., 1984.
Cobb, E.C. and Saunders, O.A., "Heat Transfer from a Rotating Disk," Proceedings,
Royal Society, Volume 236, 1956, pp. 343-351.
Dibelius, G.H. and Heinen, M., "Heat Transfer from a Rotating Disc," Proceedings of the
Gas Turbine and Aeroengine Congress and Exposition, Brussels, 1990.
Issa, R.I., "Solution of the Implicit Discretized Fluid Flow Equations by Operator
Splitting," J. Computational Physics, Vol. 62, p. 40, 1985.
Han, T., "A Navier-Stokes Analysis of Three-Dimensional Turbulent Flows Around a
Bluff Body in Ground Proximity," AIAA Paper 88-3766, 1988.
Karki, K.C. and Patankar, S. V., "Pressure Based Calculation Procedure for Viscous
Flows at All Speeds in Arbitrary Configurations," AIAA J., Vol. 27, No. 9, 1989, pp.
1167-1174.
Kershaw, D.S., "The Incomplete Cholesky - Conjugate Gradient Method for the Iterative
Solution of Systems of Linear Equations," /. Computational Physics, Vol. 26, p. 43,
1978.
Kreith, F., Taylor, J.H. and Chong, J.P., "Heat and Mass Transfer from a Rotating
Disk," Journal of Heat Transfer, Volume 81, 1959, pp. 95-105.
Kreith, F., Doughman, E. and Kozlowski, H., "Mass and Heat Transfer from an Enclosed
Rotating Disk with and without Source Flow," Journal of Heat Transfer, Volume 85,
1963, pp. 163-163.
Maliska, C.R. and Raithby, G.D., "A Method for Computing Three-Dimensional Flows
Using Non-Orthogonal Boundary-Fitted Coordinates," bit. J. of Hum. Meth. in Fluids,
Vol. 4, p. 518, 1984.
76-13
Metzger, D.E., "Heat Transfer and Pumping on a Rotating Disk with Freely Induced and
Forced Cooling," Journal of Engineering for Power ^ Volume 92, 1970, pp. 342-348.
Metzger, D.E., Mathis, W.J. and Grochowsky, L.D., "Jet Cooling at the Rim of a
Rotating Disk," Journal of Engineering for Power, Volume 101, 1979, pp. 68-72.
Metzger, D.E., Bunker, R.S. and Bosch, G., "Transient Liquid Crystal Measurement of
Local Heat Transfer on a Rotating Disk with Jet Impingement," Proceedings of Gas
Turbine and Aeroengine Congress and Exposition, Toronto, 1989.
Millsaps, K. and Pohlhausen, K., "Heat Transfer by Laminar Flow from a Rotating Plate,"
Journal of Aeronautical Science, Volume 19, 1952, pp. 120-126.
Ong, C.L. Computation of Fluid Flow and Heat Transfer in Rotating Disc Systems,
Ph.D. Thesis, University of Sussex, 1988.
Owen, j.M. and Roger.s, R.H., Flow and Heat Transfer in Rotating-Disc Systems,
Research Studies Press Ltd., Somerset, England, 1989, pp. 65-68, 93-124 and 147-149.
Parameswaran, S, and Sun, R., "Numerical Aerodynamics Simulation of Turbulent Flows
Around a Car-like Body Using the Non-Staggered Grid System," AIAA 89-1884.
Patankar, S.V., ".A Calculation Procedure for Two-Dimensional Elliptic Situations,"
Numerical Heat Transfer, Vol. 4, 1981, pp. 409-425.
Peric, M., Kessler, R. and Scheuerer, G., "Comparison of Finite Volume Numerical
Methods with Staggered and Colocated Grids," Computers and Fluids, Vol. 16, p. 389,
1988.
Popiel, C.O. and boguslawski, L., "Local Heat Transfer from a Rotating Disk in an
Impinging Round Jet," Journal of Heat Transfer, Volume 108, 1986a, pp. 357-364.
Popiel, C.O. and Boguslawski, L., "Badanie Procesow Konwekcji Masy i Ciepla Metoda
Sublimujacego Naftalenu," 1986b.
76-14
Phadke, U.P. and Owen, J.M., "An Investigation of Ingress for a 'Air-Cooled' Shrouded
Rotating Disk System with Radial-Clearance Seals," Journal of Engineering for Power,
Volume 105, 1983, pp. 178-183.
Rhie, C.M. and Chou, W.L., "Numerical Study of the Turbulent Flow Past an Airfoil with
Trailing Edge Separation," AlAA J., Vol. 21, p. 1525, 1983.
Sparrow, E.M. and Gregg, J.L., "Mass Transfer, Flow and Heat Transfer about a
Rotating Disk," Journal of Heat Transfer, Volume 82, 1960, pp. 294-302.
Van Doormaal, J.P. and Raithby, G.D., "An Evaluation of the Segregated Approach for
Predicting Incompressible Fluid Flows," ASME 85-HT-9.
76-15
1990 USAF-UES SUMMER FACULTY RESEARCH PROGRAM
GRADUATE STUDENT RESEARCH PROGRAM
Sponsored by the
AIR FORCE OFFICE OF SCIENTIFIC RESEARCH
Conducted by the
Universal Energy Systems, Inc.
FINAL REPORT
Thin Film Behavior of Povrder Lubricants
Mixed wlUi Ethylene GIvcol
Prepared by:
Academic Rank:
i
Department and
University:
Research Location:
USAF Researcher:
Date:
Contract No:
Don W. Itarelng, P.E, Ph.D.
Professor
Mechanical Engineering
University of Florida
WRDC/POSL
Wright-Patterson AFB
Dayton OH 45433-6563
Ronald D. Dayton
Aug 90
F49620-83-00053
THIN nun BEHAVIOR OF POWDER LUBRICAMTS
MIXED Wrm ETHYLENE GLYCOL
by
Don W. Dareing
ABSTRACT
The rheological properties of two powder lubricants mixed separately with ethylene
giycol were determined experimentally. The two powder lubricants ware graphite
and molybdenum disulfide. Bingham plastic, power law and hyperbolic rheological
models were compared with the experimentally determined shear stress-shear rate
rheological data. Ail three models compare favorably with experimental data at
high shear rates. The power law and hyperbolic rheological models are realistic
candidates at low shear rates. The three rheological models were incorporated Into
fluid flow equations and solved for the case of pressure induced flow between
parallel surfaces. Predicted flow rates of these non Newtonian slurries were
checked against laboratory data which was obtained through a test rig designed
and built during the study.
77-2
ACKNOWLEDGEMENTS
i wish to thank the Air Force Systems Command for giving me the opportunity to
conduct research at Wright-Patterson. I appreciate the kindness and courtesy at
aii ieveis of activity, it was an honor to associate with the miiitary professionais
and a pieasure to watch our fine young aduits in action on the Base.
i wish to thank the Air Force Office of Scientific Research for sponsoring the
Summer Facuity Research Program. The program was especiaiiy beneficiai to me
as i have recentiy transferred to the university from industry. The program heiped
me focus on specific research that is of interest to the Air Force and that has the
potentiai for iong term funding.
i wouid iike tc thank Mr Don Campbeii, Laboratory Director, for aiiowing me to be
a part of the propuision research program during the summer. Mr Benito Botteri,
POS (Chief, Fueis and Lubrication Division) and Mr Howard Jones, POSL (Chief,
Lubrication Branch) made me feei a part of their research team.
i am especiaiiy gratefui for the direct invoivement of Mr Ronaid D. Dayton, POSL
(TAM, Lubrication Branch), in the research. Mr Dayton suggested severai research
probiems to consider during my pre-summer visit, provided iaboratory space and
equipment, and supported the project throughout the summer. In addition, he
77-3
made significant technicai input to the research through periodic discussion of the
work.
Mr Nelson Forster and Mr John Schrand took time from their projects to roundup
equipment and instrumentation so that I might achieve my project goals during the
ten*week summer period. For this support, i am most grateful.
77-4
INTRODUCTION
Tribology and lubrication of machine elements are an Important consideration In
the design and operation of machinery. Equipment literally "wears out" from lack
of proper lubrication. I have been Interested In this area of technology since
conducting research on the lubrication of gear teeth for my Ph.D. (1962) at the
University of Illinois. I have written several papers on lubrication films and have
conducted research at Cambridge University on the subject of squeeze film
damping as a visiting researcher. Two summer assignments with Pratt & Whitney
dealt with structural stiffness of rolling contact bearings and traction coefficients
of transfer films In contact with soft metal coatings. Also, I teach a graduate level
course In lubrication of bearings at the University of Florida.
The Summer Faculty Research Program has given me the opportunity to become
directly involved in tribological research that is of interest to the Air Force.
Through the suggestion of Mr Ron Dayton, I chose to work on powder lubricant
slurries. This area of research Is unique in that the slurries are non Newtonian.
The summer effort focused on measuring the rheological behavior of powder
slurries, evaluating different rheological math models, and developing fluid flow
equations for pressure induced flow between parallel surfaces. The long term
application of the research is the integrated high Performance Turbi ^ Engine
Technology (IHPTET) program.
77-5
OBJECTIVES OF THE RESEARCH EFFORT
The objective of the research was to determine the thin film velocity and flow
characteristics of two powder lubricants mixed separately with ethylene glycol.
This goal is an initial step toward formulating equations of lubrication which
include the effects of these non Newtonian slurries. Steps taken during the
summer research program were:
A. Establish the rheological properties of two lubricant slurries through
testing.
B. Evaluate mathematical rheological models against measured data.
C. Develop fluid flow equations for predicting velocit/ profiles and
flow rates for non Newtonian slurries.
D. Check predicted flow rates against measured flow rates using test
equipment designed and built during the summer program.
77-6
SUMMARY OF RESEARCH CONDUCTED
UNDER SUMMER FACULTY RESEARCH PROGRAM (1990)
A. Rheology Data
Shear stress versus shear rate relationships for the powder lubricant slurries was
determined experimentally by using a Haake Rotovisco RV100 viscometer. This
Instrument has the capability of measuring the shear stress-shear rate relationship
continuously from zero rate to a maximum shear rate of about 1000 sec-1
depending upon the particular sensor system used.
The viscosity of ethylene glycol was determined early in the study and the
measured value checked with the book value of 23.0 cp at 60'’F (Reference 1).
Rheology data for ethylene glycol was determined for four temperature levels.
These data provided a baseline reference for judging the effects of powder
lubricant additives on rheology. The effects of adding molybdenum disulfide
(M0S2) to ethylene glycol are shown in the test data Figure 1. The data
corresponds to a slurry mixture of one part (weight) of MoSj to one part (weight)
of ethylene glycol. This mixture ratio (1:1) was arbitrary. These data show how
shear stress changes with temperature. Note that the relationship between shear
stress and shear strain is no longer linear.
77-7
RHEOLOGY DATA
MoS2 - Ethylene Glycol (1:1)
Swot Rote sec^- 1
FIGURE 1
RHEOLOGY DATA
Graphite-Ethylene Glycol (1:8)
Siear Rota sec — ^1
FIGURE 2
77-8
The effects on rheology of adding graphite to ethylene glycol are shown in the test
data given in Figure 2. The data corresponds to a slurry mixture of one part
(weight) of graphite to eight parts (weight) of ethylene glycol. This mixture ratio
(1 :8) was also arbitrary, however, higher ratios were very thick and difficult to test.
The main characteristic of this slurry is the high degree of nonlinearity of the
curves at low shear rates. These slurries are thixotropic as they shear thin at high
shear rates.
B. Rheology Mathematical Models
Three different mathematical models were considered for approximating the
experimental rheology data. Each of these models are discussed below.
1. Bingham Model: This model (References 2,3) is mathematically the
simplest of the three because it is a straight line intersecting the shear stress axis
at Tq called the yield point.
t= To + Ko/ (1)
This model describes a substance that has no shear rate until a certain shear
stress level has been reached. This model deviates considerably from the
measured data at low shear rates as the measured data shows no yield point.
However, the model is accurate at high shear rates where many lubricating films
operate.
77-9
2. Power Law Model: This model (Reference 2) is so named because the
shear rate variable is raised to an "n th" power as defined in Equation 2. The
model is useful for describing shear stress-shear rate relationships for non
Newtonian fluids that have no yield point. This is the case here. The model is
amenable to fluid flow formulations.
T = Kf (2)
The power law model may be most useful in predicting laminar flow through
capillary tubes. This model also covers the full range of shear rates.
3. Hyperbolic Model: This model can be formulated to match the
experimental data from zero shear rate to very high shear rates. However, its
mathematical expression is more complex and thus velocity profile and flow rate
predictions are more difficult to obtain.
T= (ko;^2K,TojT (3)
This function approaches the Bingham line at high shear rates and matches
reasonably well with the experimental data at low and near zero shear rates. The
constants on the mathematical expression are the same as the ones used in the
Bingham model expression.
77-10
C. Fiuid Flow Equations
The fluid flow equations were developed for pressure induced flow between parallel
surfaces. The purpose here Is to gain insight into the behavior of these non
Newtonian lubricant slurries in thin films. Velocity induced flow formulation and
Reynolds equation follow directly from these equations.
The velocity profiles for the molybdenum disulfide and graphite slurries (Figures
3 and 4) show the effects of the nonlinearity of the shear stress-shear rate curves
(Figures 1 and 2). The conditions for these curves are also
h s 0.0005 inches
dp/dx = 20 psi/inch
Temp = 20*0
The velocity profile curves in this case are more blunt near the center of the film.
This effect would be amplified by larger values of Tq. The zone of plug flow for the
Bingham model is
dx
The extent of the plug flow zone is identified by the point "a" in the figure. The
hyperboiic modei profiie more nearly approximates the true velocity profile because
77-11
\fabcly.
VELOCITY PROFILES
MoS2 - Ethylene Glycol {1:1)
POWER
Location in Fin
FIGURE 3
VELOCITY PROFILES
Graphite-Ethylene Glycol (1:8)
Location h Fin
FIGURE 4
77-12
flaw ftjta. IV'Z/aac
this model more nearly fits the shear stress>shear rate test data at low shear rates
(0 to 1000 sec-1).
Flow rate calculations were made using the above rheological models. An example
set of calculations are shown In Figure 5.
The ethylene glycol case was determined from classical Newtonian equations of
flow (Reference 4). For the two powder slurries, all three rheological models
predict essentially the sanM flow rates. This Is due to the fact that shear rates
within the films are far In excess of the 1000 sec'' and In the operating region
where all three models converge.
Flow Rate Predictions
Film 'rhickness is 0.002 inches
ETHYLENE
GLYCOL
GRAPHITE
SLURRY
HoS,
SLURRY
0 200 400 600 800
R^essure GrocTient psi/ncfi
FIGURE 5
77-13
D. LABORATORY TESTING
A laboratory test rig was designed and built to check fluid flow predictions based
on the rheological data and equations of flow given earlier. The rig was designed
to generate pressure induced flow through parallel surfaces.
Parallel surfaces were approximated by an annular space having a circumference
of two (2) inches. The length of the inner cylinder was one inch so that pressure
gradient could be ascertained directly from pressure measurements which were
determined from a load cell.
Flow across the parallel surfaces were established by advancing a piston by means
of a power screw driven by a 1/2 horsepower motor thru a four to one timing belt
pulley pair. This arrangement provided a positive displacement pump having the
capability for low volume delivery at high pressures. After a test fluid is forced
through the parallel surfaces, the piston is reset and the cylinder rech<irged for
another test run. A tachometer gave volume flow while a load cell mounted in
tandem with the piston rod gave fiuid pressure.
The test rig aiiowed fiow rate versus pressure data to be coiiected for smaii
sampies of the powder iubricant siurries.
Data from these experiments are aiso given in Figure 5. in generai there was
reasonabiy good agreement between theory predictions of fiow and the test data.
77-14
RECOMMENDATIONS
The summer research established velocity and flow rates for pressure induced flow
through parallel surfaces for two powder lubricant slurries. This was a useful first
step, because pressure induced flow is an important component of flow in
lubrication films. The extension of this research should combine the results of
pressure induced flow with velocity induced flow to formulate equations of
lubrication. Recommended future work is outlined below:
A. Develop Reynolds equation of lubrication for rheological models which
closely match the experimental data obtained under the SFRP (1990).
B. Evaluate techniques for solving Reynolds equation for Non Newtonian fluids
including methods
1. found during extensive literature survey.
2. developed during the research
C. Solve Reynolds equation for:
1 . Hydrostatic type bearing.
2. Squeeze film type bearing.
3. Slider bearing (no side leakage).
D. Design and build test equipment to measure performance of the above three
types of bearings.
77-15
E. Compare theory predictions of fiim performance against laboratory test data.
Laboratory data should be obtained under different operating conditions for
the three different types of lubrication film.
Longer term work should focus on the behavior of powder lubricant slurries
in the elastohydrodynamic lubrication of rolling contact bearings.
77-16
REFERENCES
1. Windholz, M., The Merck Index. 10th Edition, Merck and Co., Inc., 1983.
2. Skeiland, A. H. P., Non Newtonian Flow and Heat Transfer. John Wiley and Sons,
1967.
3. Cameron, A., Principles of Lubrication. John Wiley and Sons, 1966.
4. Fuller, D. D.: Theory and Practice of Lubrication for Engineers. 2nd Edition,
John Wiley and Sons, 1984.
NOMENCLATUKS
dp/dx
h
K
Ko
n
yo
t
to
y
Preaaurm gradient
Film thickneaa
Po^ar Law conatant
Bingham model conatant
Power Law exponent
Limit of plug flow
Shear stress Pascals
Yield stress in Bingham model
Shear rate, sec'^
77-17
1990 TJSAF-UES SUMMER FACULTY RESEARCH PROGRAM/
GRADUATE STUDENT RESEARCH PROGRAM
Sponsored by the
AIR FORCE OFFICE OF SCIENTIFIC RESEARCH
Conducted by the
Universal Energy Systems, Inc.
FINAL REPORT
Lassr Vslogirngtry Mgasurgmgpts in.Shpgk Thb??
Prepared by:
Academic Rank:
Department and
University:
Research Location:
USAF Researcher:
Date:
Paul A. Dellenback
Assistant Professor
Civil and Mechanical Engineering
Southern Methodist University
WRDC/POTC
Wrignt-Patterson AFB Ohio 45433
Major Geoffrey Jumper, USAF
July 27, 1990
F49620-88-C-0053
Contract No:
by
Paul A. Dellenback
ABSTRACT
The overall objective of this research was to measure turbulence intensities in
shock tubes with non-intrusive optical instrumentation. Laser velocimetry measurements
applied to shock tube driven flows present several new problem areas that require
attention. Among these are the implementation of appropriate circuitry for triggering and
controlling high speed data acquisition, and the development of a suitable seeding
material for these flows. Control circuitry was readily developed and demonstrated, but
equipment limitations severely impeded success in resolution of particle seeding issues.
78-2
Acknowledgements
The support of the Air Force Systems Command, the Air Force Office of
Scientific Research, and the Aero Propulsion Laboratory at Wright-Patterson AFB are
gratefully acknowledged. Major Geoffrey Jumper's leadership and personal attention
were instrumental in my enjoyment of this work. Mr. Greg Gala and Mr. John Schmoll
were endlessly helpful in numerous logistic and equipment requirements that developed
during the course of the investigation.
1. Introduction
The Air Force and those companies that build its propulsion engines are
continuously involved in concentrated efforts to improve gas turbine engine performance.
For the last decade, much of their attention has focused on higher turbine operating
temperatures. Enhanced engine performance requires, among other things, a detailed
understanding of both the convective heat transfer and the flow fields associated with the
high temperature turbine components. The relationship between free stream turbulence
intensity and local convective heat transfer rates has especially been of interest in recent
years.
The overall research effort ongoing in this country to understand the interaction
between convective heat transfer and free stream turbulence levels has many different
facets. One component of that program is an in-house effort at WPAFB to directly
measure both quantities in shock tubes designed to accommodate full scale rotating
turbines. The motivation for using shock nibes is to reduce testing expenses to
manageable levels, but at the same time, several new complications are introduced
beyond those normally encountered in steady flow wind tunnel test facilities.
Identification of, examination of, and resolution of these complications has been ongoing
at WPAFB, and the purpose of this SFRP effort was to both contribute to this
development effort and to initiate actual measurements of turbulence intensity in a full
scale shock tube facility at Calspan Corp. in Buffalo, NY.
The author's possesses a solid background in the field of optical measurements as
applied to turbulent flows, while at the same time, the author's principle research interests
are in experimental convective heat transfer. Thus, there was an ideal match between the
author's qualifications and the problem needs associated with the problem discussed
above. Major Geoffrey Jumper, USAF, recognized this match and hosted the author's
studies at WPAFB to assist in the shock tube velocimetry investigations.
78-4
11. Objectives of the Research Effort
The overall research objectives as outlined by Major Geoffrey Jumper included;
a) becoming familiar with use of POTC lab’s laser two-focus (L2F) velocimetry
system,
b) developing, in conjunction with technical support from Mr. John Schmoll,
suitable triggering circuitry for use in a shock tube,
c) forwarding the work of Rivir, Elrod, and Dunn^ by examining flow seeding
alternatives suitable for the high temperature (1000 deg F) environments
that were anticipated in Calspan's full scale shock tube,
d) travelling to Buffalo, NY for a two-three week period for the purpose of
performing a series of measurements in Calspan Corp.'s full scale turbine
test rig (a large scale shock tube),
e) returning to WPAFB for final data manipulation and writing up of findings
and results.
The above series of events were viewed, even from the outset, as an ambitious
program for a ten week time frame by both Major Jumper and the author. Unfortunately,
equipment limitations severely compromised the overall success of achieving the
objectives stated above.
III. Experiences with the L2F System
The only aspect of the laser two-focus (L2F) velocimeter familiar to the author at
the outset of the project was the basic operating principle whereby two parallel laser
beams separated by a known distance are used to monitor the time-of-flight associated
with micron-sized particles passing through one beam and then the other. The particle
velocity is then computed by dividing the beam spacing by the time-of-flight. The
inherent advantages associated with L2F techniques include the ability to measure very
near walls and generally superior signal-to-noise ratios (SNRs) compared with more
conventional laser Doppler anemometry. However, the L2F measurement is very
directionally sensitive so that only particles traveling through a very narrow band of
angles will intercept both beams. The L2F technique thus has historically been limited
78-5
primarily to flows that have a single dominant flow direction and to those having fairly
low levels of turbulence. The L2F is limited to these applications in order to assure that a
single particle crosses both beams, as opposed to the ambiguous case where one particle
passes through the first beam and a second particle passes through the second beam. This
latter series of events becomes more probable for multi-dimensional or highly turbulent
flows.
Familiarization with the basic Polytec L2F velocimeter occurred during the first
week of the program. The shock tube facility to be used in support of the above-stated
seeding and firing circuit development was located in the Air Force Institute of
Technology (AFIT), and due to its primary role of teaching support, was unavailable
during the initial three weeks of the project. Hence, initial investigations into L2F
operations were conducted in a small (1 cm diameter) free jet seeded with droplets of
propolene glycol.
Two problems were encountered during early L2F usage. The particular
instrument used here contained two photomultiplier tubes (PMTs), with each monitoring
the waist region (i.e. smallest diameter region) of a single beam. Thus, outputs from the
two beams can be, and in fact must be, simultaneously monitored for alignment purposes
as well as for correctly adjusting the controls of the dedicated L2F signal processor. Dr.
R. Rivir, the last user of the L2F instrument, reported before the project began that one of
the two PMTs was misaligned so that it was not obtaining any signal from the beam it
was intended to monitor. However, with some good fortune, both PMTs were initially in
alignment so that scattering signals were in fact being received. Unfortunately though,
one of the two signals was oddly distorted so that rather than appearing gaussian as
expected, it appeared somewhat like a cresting wave. This shape clearly would
negatively impact the overall accuracy of the L2F measurements (albeit to a small
degree) since the pulse shaper and discriminator in the signal processing electronics
would be "confused" by the peculiar slope of the signal, thus producing a systematic
78-6
timing error. By swapping the two PMT’s and their associated support circuitry in
various ways, it was detemiined that the source of the misshapen signal was the receiving
optics alignment. By disassembling the receiving optics and realignment, the problem
was eventually resolved.
A second obstacle to immediate success was more readily managed. The rather
complicated software supplied by Polytec is intended to be used in steady flows in which
the plane containing the two laser beams can be leisurely rotated through 90 degrees to
allow determination of the predominant flow direction. Unfortunately, the software will
only work in this mode. Clearly then, the software is not satisfactory for transferring data
from the dedicated L2F electronics and processing for the shock tube measurements
because no such rotation of beam orientation is possible in the 10-40 msec time window
for data acquisition. A scheme to "fool" the software by supplying arbitrary and
artificially generated angular data was conceived and tested to assure that it produced no
adverse effects on accurate data reduction. This scheme was determined as sound and
subsequently used to reduce one-dimensional shock tube data. It perhaps should be noted
that a data reduction program had previously been written by a former student for use
with the L2F system, but critical inspection of this routine by Mr. Schmoll and the author
established that there were at least eight conceptual and algorithmic errors in the code so
that its use was abandoned as unfeasible without a substantial effort towards revision.
It should also be noted that neither data reduction program discussed above
appears to incorporate a correction scheme for the velocity bias problem associated with
finite-sized probe volumes, even though by the mid-1980s this effect has been
established as a non-trivial contributor to errors in calculation of mean velocity and
turbulence intensity^.
The beam diameters at the waists and the beam spacing were measured as per the
Polytec instruction manual by projecting the measurement volume through a graduated
microscopic slide and a microscope objective lens. In this way, the beam spacing was
78-7
determined to be 350 microns and the beam diameter was determined to be about 12
microns.
The L2F does not lend itself to an easy determination of data rate since pairs of
signals received by the PMTs must be correlated with many other pairs to determine the
likelihood that a single particle has produced both scattering events. Thus, a data
acquisition rate can only be determined at the end of a timed period by counting the
number of properly correlated events. However, it is a straightforward matter to count
the particle crossings seen by each PMT, and the Polytec L2F instrument provides such
an on-line readout of "event rates". The event rates will always be larger than the data
rate. The data rates are of critical interest in the present study since it is desired to
perform a statistical analysis on the data (i.e. to determine the RMS fluctuation of
velocity, that when normalized is referred to as "turbulence intensity"). Furthermore, the
error associated with the sample size is inversely proportional to the square root of the
number of samples, so it is desirable to obtain several thousan 1 such data points per
location. It is somewhat tricky to exactly reproduce the conditions in a shock tube from
shot to shot, so that it is further desirable to obtain several thousand data points in any
one 20 to 40 msec shot. Assuming a minimum number of samples is 4000, then in a 20
msec shot (which is actually a bit long for the AFIT shock tube), one must have a data
rate of at least 200,000 Hz, and event rates that are larger still. Thus, mention of event
rates will occur in the discussion which follows, and in fact the search for very high data
rates was the motivation for examining the effects of system gain and seed density that
are discussed in the following paragraph.
The impact of seed density was investigated in the free jet. In a series of tests it
was determined that although signal noise is a strong function of PMT voltage (wliich is
like an amplifier gain in the system), the accuracy of computed mean velocity and
turbulence intensities (TIs) does not appear to be negatively impacted by even very high
gains that result in noisy signals. A similar systematic investigation of various seed
78-8
densities (employing propolene glycol) in the free jet likewise showed little impact on the
computed velocity statistics.
The event rates in the PMT gain and seed density studies discussed above were
between about 1000 and 25,000 Hz on each channel of the L2F optics. While low by the
criteria stated above, velocities in the free jet were about one order of magnitude lower
than behind the M=1.3 shocks that were tj^pically generated in the AFIT chock tube
facility. Because theoretical data rate is proportional to jet velocity, or how fast seed
droplets are swept through the measurement space, the observed event rate of 25,000 Hz
in the free jet study suggests that event rates of 250,000 Hz might be possible in the AFIT
shock tube if the flow can be seeded to a similar density. It should be noted that in the
free jet studies, an event rate of 25,000 Hz was produced ai the expense of very noisy
signals and it was surprising to the author that the ensuing mean and RMS velocities
agreed so well with cleaner, less noisy signals obtained at lower seed densities and lower
system gains.
Eventually, the AFIT shock tube facility became available and was reconfigured
to provide a test region one inch square. The tube was essentially set up as it was when
Rivir, et Al.^ made L2F measurements during the 1983/84 time frame.
The tube was first operated in a continuous flow mode to allow for determination
of beam angle. With a mean velocity of 120 m/s (the largest available from the air
supply), the impact of L2F angular bias on accuracy of mean and RMS velocities was
investigated. The plane containing the beams was tilted about the mean flow direction a
total of plus and minus 6 degrees, in increments of two degrees (see Figure 1). When the
plane of the beams was not aligned with the flow direction, the indicated mean velocity
was too low by about 4% per degree of beam-plane rotation. The sense and magnitude of
this bias error in mean velocity are dependant on the seed density, but the magnitude of
the 4^c figure is interesting as it pertains to what seemed to be fairly typical measurement
conditions. The TI tended to increase with beam angle error at about 6% per degree of
78-9
beam-plane rotation over the range examined here (for TIs computed based on the actual
mean velocity of 120 m/s). The TI results are qualitatively as expected, and demonstrate
the potential for biasing of data due to the directionally sensitive L2F measurement
space.
With this operational experience, our attentioji turned to making measurements
during 10 msec windows behind shock waves. The window duration of 10 msec was
chosen after inspecting transient, representative shock pressure histories produced by a
pressure transducer located just upstream of the optical window access and in the one
inch square cross section. This period is only half that used previously by Rivir, et. Al,
but it was felt that our choice would provide the more accurate measure of true
turbulence intensity as the pressure (and presumably the flow velocity) is clearly
decreasing for all time after passage of the shock. It has been assumed that the data rates
obtained in our measurements could be roughly doubled if the window for accumulating
data were increased to 20 msec. The strength of the shocks for all runs was about 1.3, as
determined from calculations for an ideal shock tube based on the material in Gaydon^.
Towards the end of a series of roughly 60 experience-generating tube shots, a
careful set of ten shots were contrasted. For these cases, computed mean velocities were
consistently in the 245 m/s regime (plus or minus about 5 m/s). However, TI were less
consistent as they were observed to vary between 3.3% and 8.4%, depending on the many
operator-settable parameters employed. The inconsistency of the TI measurements and
our inability to determine which subtle influences were altering these values from run to
run were very disturbing. Furthermore, the number of satisfactorily correlated data that
were obtained in any 10 msec window were typically 100 to 200 velocity measurements.
This, based on the preceding discussion, was also very disappointing since the sample
size was generally too small for a satisfactory statistical analysis.
It was felt that development efforts should proceed with a seed media that was
more like that we hoped to implement in the full scale Calspan tests that would require a
78-10
seed material with higher temperature capability than the propolene glycol. Solid particle
seeds were investigated because they tend to be convenient to use (e.g. as opposed to
titanium tetrachloride), but there may ultimately be some concern as to their abrasiveness
in the Calspan tunnel. A small swirling-flow seeder was installed in place of the six-jet
atomizer. Micro-balloons particles were used initially due their well-behaved nature and
ready availability (that is, micro-balloons shows a minimal tendency to clump relative to
many other common solid particle seed materials and they were given to us by a UD
employee). During set-up of the solid-particle seeder with continuous flow through the
shock tube, it was clear that a problem had developed.
Use of the solid seed generally required higher PMT voltages due to the smaller
size of the new seed. As we attempted operation at the required PMT voltages, fairly
regular 0.5 volt "events" spaced at approximately 5 micro-second intervals were in
obvious profusion on one channel of the L2F. This noise was not related to seeding, as
ascertained by turning off the seeder. Eventually, the source of the noise was determined
to be the PMT itself. This caused the L2F work to come to a halt since the nature of the
noise was such that it looked like scattered light from particles passing with a frequency
of 250 kHz was being accumulated by the PMT. Tlie L2F electronics was all too happy
to process these signals. A replacen.ent PMT was ordered, but tlie quoted delivery period
(4 to 6 weeks) was so extensive that alternative approaches to the problem at hand were
clearly in need of consideration, given the short 10 week period available to complete the
work.
IV. Experiences With the LDA
A laser Doppler anemometer (LDA) was available for use in these experiments.
The author's background in LDA measurements and the technique's general superiority
over the L2F method for making measurements of TI led to optimism that it could be
78-11
usefully employed on the present problem. It was also anticipated that the replacement
PMT for the L2F might arrive towards the end of the investigator's 10 week tenure so that
simultaneous, side-by-side L2F and LDA measurements might be conducted for purposes
of comparison and confidence building in the L2Fs TT measurement capability. It was
recognized that the LDA's inherently lower SNRs might negatively impact the
instrument's performance in this environment where backscatter light collection is
required.
A plan was formulated whereby it would be attempted to install the LDA into the
AFIT lab and make velocity measurements. Two logistic problems associated with
actually using the LDA presented themselves. First, there was not a proven data
acquisition system and interface available to work with the IBM XT computer that was
available to support these experiments. Secondly, simultaneous measurements with both
the L2F and LDA seemed to require a physically complex arrangement for support and
mounting of the two systems. The second problem was to be managed by incorporating a
fiber optic probe with the LDA, thus allowing remote placement of the laser and optics.
The penalty to be paid was that the fiber optic probe was of a rather old design such that
only very low light transmission could be obtained.
The problem associated with hardware interfaces, software interface drivers, and
data acquisition and processing software was capably tackled by Mr. Schmoll, who had
previously generated a substantial amount of code for exactly these purposes.
Modifications to Mr. Schmoll's routine required about one week, but this step was
important to determining the viability of LDA use in this investigation since there was
concern about the maximum data through-put capability of the LDA processor and
computer interface. Through-put rate is not a concern in the L2F instrument where data
points are accumulated in real time by a multi-channel analyzer and then later dumped to
the computer for processing at the user's leisure. Eventually, Mr. Schmoll's program
development was sufficiently advanced that we could measure the through-put rate.
78-12
However, the software was never completed due to Mr Schmoll's efforts being required
on other projects, and thus no capability of storing data and statistically processing data
was available, 'fhe data through-put rate of the TSI interface was determined to be
marginally adequate, but no single quantitative value was determined since the rate
proved to be a function of the validated data rate determined by the counter. Generally
though, the through-put rate was approximately 80% of the validated data rate for low
data rates, and as little as 30% for the highest data rates examined. We guessed that in
optimal conditions, a thousand or so data points could be accumulated in our 10 msec
window, and if this were in fact the case, it would represent a substantial improvement
over our prior experience with the L2F system.
Based on the through-put measurements, the decision to move and try the LDA
was made. A delay of almost two weeks ensued as power for the LDA's laser was
installed in the AFIT lab. Note that the L2F's laser and the LDA's laser had substantially
different power requirements (single phase vs. three phase, respectively), thus requiring
the new electrical service.
The LDA was installed and aligned. Alignment proved very tedious with that
portion of the system which "piped" the laser beams into the optical fibers. With a beam
collimator, two mirrors, a beamsplitter, Bragg cell, and fiber-optic link installed in the
optical train, the best optical transmission efficiencies for the entire system were about
5%, thus providing a maximum of 45 mW of optical energy at the probe volume
(employing the 1 watt, 488 nm line from the argon laser). When contrasted with the 200
mW available at the probe volume from the L2F system and the smaller laser beam foci
associated with the latter, it is clear that the L2F system had a substantial advantage in
potential SNR that could be achieved (a rough computation suggests that the SNR under
conditions of these tests might be as much as 1900 times higher with the L2F). The LDA
probe volume could not be successfully imaged back through the fiber-optic probe, even
under ideal fixed- scattering source conditions. In tracking down this apparent alignment
78-13
problem, the fiber at the PMT head was either discovered to be severed, or was severed
as a consequence of our investigations. The fiber was not repairable at that time due to a
lack of tools and experience in such matters. Thus the probe volume was imaged via a
standard optical housing and PMT mounting assembly. Light collection at the front end
of the receiving optics was improvised with a lens having focal length of about 100 mm
and aperture of 40 mm.
The make-shift LDA proved successful in operation with propolene-glycol
seeded, low velocity flows. However, success with the LDA configuration used here in
shock tube measurements was minimal.
One persistent problem precluded success with high velocity steady flow
measurements. Shot noise from the PMT could be observed on an oscilloscope as a
single, classic, spike. The PMT signals were processed in a dedicand LDA counter that
was typical in that its front end contained an amplifier and bandpass filters. The
amplifier apparently introduces substantial overshoot into the incoming signal so that the
PMT shot noise was modified. After passing through the filtering and amplification
stages of the counter, observation of the shot noise signal on an oscilloscope showed that
it contained a number of distinct cycles. This signal contained an even greater number of
cycles if the PMT output were first passed to the frequency downmixing circuitry of the
Bragg cell power supply and then to the counter, as is the usual practice. Thus, the shot
noise was being modified by the processing electronics to appear exactly like a Doppler
burst! Not surprisingly, the timing circuits in the counter were incapable of detecting a
difference between the artificially generated shot noise "bursts" and true velocity samples
(note that the shot noise problem occurred even when the PMT aperture was masked off
from all light sources). Apparent data rates due just to the shot noise contribution were
measured at 330,000 Hz when measuring the time for 8 cycles per burst. Increasing the
number of cycles per burst to 16 only halved this data rate - that is, the modification to
the original shot noise signal contained at least 16 cycles. Counting 32 cycles per burst
78-14
was similarly a problem, but judged not feasible due to the calculated number of fringes
(50) in the probe volume. Filtering of these signals was also of no use since the apparent
frequencies were about 20 MHz, which fell in the range expected for steady flow
measurements. For measurements behind shock waves, the combination of expected
velocities (determined as about 250 m/s from L2F measurements), fringe spacing (4.89
microns), and possible frequency shifts suggested that the shot noise generated signals
could be filtered out by passing those signals with frequency content greater than 50
MHz, and this could indeed be accomplished with the filter circuitry of tlie counter.
The final conclusions pertaining to the shot noise problem should be noted by
those who intend to do LDA work at higher velocities and therefore higher frequencies:
The TSI counter processor takes a noise signal that a user has every expectation will be
ignored, and modifies it to appear as a plausible Doppler burst, which the counter is then
happy to process. Obviously, there will be a severe biasing of data if this problem is
unaccounted for. It would be interesting to see how the counter reacts to a clean pulse
generated by a pulse generator. As an editorial note, it is remarkable what junk one may
be sold for a $20,000 outlay (i.e the price of a TSI counter).
The shot noise problem fooled this author for some time. When discovered and
filtered out for the shot tube tests, the LDA system used here proved :o be inadequate for
the job. An event counter was set up to record the number of samples during a 10 msec
window behind the shock wave, but results from seemingly identical mns were so
inconsistent that no conclusions could be reached. For several shots, several hundred
data points were apparently generated, while for most runs, no bursts with frequencies
above 50 MHz were recorded. Captured PMT signals on an oscilloscope suggested that
the electronics was worxing correctly, as there were no discemable bursts in the
background noise. Because they could not be repeated, those runs for which several
hundred data samples were accumulated are very suspicious.
78-15
In conclusion, it appears that SNRs were simply too low for detection and
processing by the counter. Note that even a ten-fold increase in probe volume intensity
that could be expected from eliminating the fiber-optic link (or by employing a state-of-
the-art fiber optic system) might not be sufficient to give satisfactory results. This
observation can be put in perspective by noting that the SNRs estimated from
oscilloscope observations of the L2F signals were often less than 10, and even after LDA
system improvements, the LDA probe volume intensities (and by inference, the SNRs)
would still be 190 times lower than for the L2F.
Low SNRs in these experiments were further aggravated by the necessity to use
off-axis backscatter light collection. An attempt to use forward-scatter light collection
was made, but the requirement to remove and clean the shock tube windows after about
every 4 shots, coupled with the short focal lengths associated with the optics available to
the author, rendered this effectively impossible. Specifically, the forward-scatter optics
had to be realigned after every window cleaning. Realignment required a window¬
dirtying flow with scattering source, so that the number of shots until the next cleaning
and realignment was reduced to about one or two. Any future effort to use an LDA for
verifying L2F accuracy should employ forward scatter optics with sufficiently long focal
lengths so that realignment is not required after each window removal and cleaning. The
forward scatter optics in combination with removal of the fiber-optic link should provide
SNRs on the same order as those available from the L2F velocimeter (i.e., SNRs could be
expected to be at least 1000 times higher than available with the optical arrangement
I employed here). Alignment of receiving optics also proved to be tedious due to the
make-shift receiving lens arrangement, scattered light from the rear test section wall (or
window in the case of forward scatter), and to an unusually bright background glow of
laser light in the test region.
78-16
V. Conclusions and Recommendations
It is clear that performing measurements in shock tube driven flows adds
complication over and above that encountered in steady flows. With respect to the
original objectives of the investigation, it can be noted that Mr. Schmoll built a timing
and gating circuit that worked satisfactorily throughout the latter eight weeks of the
study. However, the effort this summer became so bogged down in various other areas
that the issue of a suitable seed material is yet unresolved.
A problem not addressed during the brief periods of testing v/ith a solid particle
seed material, but one by no means trivial, was the abrasiveness associated with the
particle's high speed impact with surfaces. Although the particles are small (on the order
of 0.5 to 1 micron), they are on the same order as the size of polishing grit employed to
fabricate thin-film heat flux gauges on the turbines tested at Calspan. Hence, it is not
clear that solid particle seeding, even if optimal for the optical measurements, would be
acceptable to use in the Calspan tunnel or other tunnels due to possible abrasive action on
other instrumentation.
This author's concerns associated with the accuracy of L2F techniques for
measurement of T1 are also unresolved. It is this author's opinion that any effort made
specifically for the purpose of measuring TI, and in which the principle instrument used
is an L2F, will be viewed as suspicious by the experimental turbulence and heat transfer
communities. Thus, I would urge a renewed effort to collaborate the L2F's ability to
accurately measure TI in this difficult flow measurement problem through side-by-side
comparison with either LDA or hot-wire measurements. As shock tubes become more
common for use in what were traditionally steady flow investigations (e.g. WPAFB's new
ATARR facility), it will be important to establish early on the reliability of one's
measurement technique.
78-17
Any future efforts to verify L2F operation with an LDA should employ forward
scatter measurements to insure SNRs on the same order as those available with the L2F
instrument. (It is recognized that forward-scatter optics can not be used in many actual
measurements, as for example at Calspan, but it could be useful for accuracy verification
in the AFIT shock tube.)
One difficulty that may arise in attempting to contrast simultaneous hot-wire and
optical measurement methods is that hot-wires may not be sturdy enough to withstand
bombardment by the seed particles. There is also the possibility of using the hot-wire
simultaneously in a liquid droplet-seeded flow, but there will be a problem with liquid
accumulation on the probe and possible "bombardment breakage" of the wire (?). There
is some local experience^ with using hot-wire techniques to measure turbulence
intensities in shock tube driven flows that could be helpful to future efforts.
78-18
References
1) Rivir, R.B., M.G. Dunn, and W.C. Elrod, "Feasibility Study of the Application of a
Two-Spot Laser Velocimeter to Measure Velocity and Turbulence in a
Shock Tube and in a Turbine Stage", AFWAL-TR-84-2054, 1984.
2) Stevenson, W. H., H.D. Thompson, and T.C. Roesler, "Direct Measurement of Laser
Velocimetry Bias Errors in Turbulent Flow", AIAA L, 2Q, pp. 1720-1723,
1982.
3) Gaydon, A.G. and I.R. Hurle, The Shock Tube in High Temperature Chemical
Physics. Reinhold Pub. Co., 1963.
4) Rockwell, R., P. King, and W. Elrod, "An Electrical Analog Circuit for Heat Transfer
Measurements on a Flat Plate Simulating Turbine Vane Heat Transfer in
Turbulent Flow", AIAA Paper no. 90-2412, 1990.
78-19
120
78-20
Figure 1 - Impact of L2F beam angle orientation errors on mean velocity
and turbulence intensity in a steady, 1-D flow
1990 USAF-UES SUMMER FACULTY RESEARCH PROGRAM/
GRADUATE STUDENT RESEARCH PROGRAM
Sponsored by the
AIR FORCE OFFICE OF SCIENTIFIC RESEARCH
Conducted by the
Universal Energy Systems, Inc.
FINAL REPORT
Thermal Analysis of Potential Solid Lubricant Candidates
Prepared by:
Academic Rank:
Department and
University:
Research Location:
USAF Researcher:
Date:
Contract No:
Dennis R. Flentge, Ph. D.
Associate Professor
Science and Mathematics
Cedarville College
WRDC/POSL
Wright Patterson AFB
Dayton, OH
Phillip Centers, Ph. D.
September 14, 1990
F49620-88-C-0053
Thermal Analysis of Potential Solid Lubricant Candidates
by
Dennis R. Flentge, Ph. D.
ABSTRACT
A series of tungsten, molybdenum, phosphorus, and zirconium
compounds were studied using thermogravimetric analysis, differential
thermal analysis, and mass spectrometry. Interaction of these materials
with silicon carbide and silicon nitride were also examined. Some
evidence was found for the catalytic effect of lead cations on the
conversion of the carbide and nitride to silicon dioxide. Oxythiomolyb-
dates and oxythiotungstates released sulfur dioxide when heated and
promoted the conversion of silicon carbide and silicon nitride to silicon
dioxide.
79-2
ACKNOWLEDGEMENTS
I wish to thank Air Forces Systems Command and the Air Force Office
of Scientific Research for sponsorship of this research. I also wish to
thank Universal Energy Systems, Inc. for their administrative assistance
for the progrcun.
Interactions with a number of scientists in the lubrication
laboratories made the laboratory work enjoyable and successful. Dr.
Phillip Centers provided guidance and encouragement. Bob Wright and Capt.
Charles Kelley gave both technical advice and scientific counsel. Chris
Klenke, Lynne Nelson, Rita Chan, Al Beane, and others provided personal
support that made the summer experience a pleasant one.
79-3
I . INTRODUCTION
The lubrication laboratory in the Propulsion Laboratory of the
Wright Research and Development Center at Wright Patterson AFB has been
investigating a series of compounds which could be used as high tempera¬
ture lubricants. During the past six years I have studied the degradation
of polyphenyl ethers (a class of liquid lubricants) and performed some
preliminary studies on antimony sulfides (a class of solid lubricants).
My graduate studies were in the area of surface catalysis. This
background prepared me well for the research conducted this summer.
II. OBJECTIVES OF THE RESEARCH
The United States Air Force has a research goal of substantially
increasing the output of its turbine engines during the next 15 years.
This increase in output will be accompanied by an increase in operating
temperature which will place an increased demand on the performance of the
lubricant being used. Several solids are under consideration as potential
lubricant candidates or as precursors of acceptable candidates. Since
molybdenum compounds have shown good lubricating capabilities at lower
temperatures several molybdenum compounds and their analogues are being
considered for the higher temperature use. In this research a series of
molybdates, oxythiomolybdates, tungstates, oxythiotungstates, phosphates,
and zirconates were studies using thermogravimetric analysis, differential
thermal analysis, and mass spectrometry.
Our primary goals were to (1) determine the thermal and oxidative
behavior of the molybdates, oxythiomolybdates, and analogous compounds,
and (2) determine the nature of the chemical and physical reactions that
occur between the surfaces of the these materials and silicon carbide or
silicon nitride.
79-4
III. EXPERIMENTAL
The thermal analysis data were collected using the DuPont Thermal
Analyst 2100 System with the 951 Thermogravimetric Analyzer and the 910
Differential Scanning Calorimeter. Gas phase products generated during
the thermogravimetric analysis were analyzed using the VG Quadrupoles
Micromass PC.
Lead tungstate, lead molybdate, lead phosphate, cadmium tungstate,
cadmium molybdate, zinc molybdate, and zinc zirconate were synthesized by
Comprehensive Research Chemical Corporation. Cesium oxytrithiomolybdate
was synthesized by Pennwalt Chemical Corporation. The remaining materials
were synthesized for the lubrication laboratory under an Air Force
contract .
Mixtures of silicon nitride (Si]N4) or silicon carbide (SiC) with the
substances studied were prepared in a 1:1 mass ratio of substance to
nitride or carbide.
IV. RESULTS
The compounds studied can be grouped into two categories — compounds
containing no sulfur and oxythiomolybdates or oxythiotungstates.
A. SULFUR-FREE COMPOUNDS
Table I contains the thermal and mass spectral data for the sulfur
free compounds. Cadmium tungstate, magnesium tungstate and lead tungstate
show little mass loss when heated. Only the lead tungstate shows a mass
gain when SiC or Si3N4 were added to the pure material. These mass gains
are accompanied by the liberation of heat and generation of carbon dioxide
(COj) or nitrogen monoxide (NO).
Table I Tungstates and Analogous Compounds — Ther¬
mal and Mass Spectral Data
T6A
mass %
change*
1) CdW04 <1
w/SiC +3
w/SijN^ 1.5
2) MgWO^ <1
w/SiC +3
w/Si3N4 +1.5
3 ) PbW04 <1
w/SiC +15
w/Si3N4 +12.5
4) Pb3(P04)j -7.5
w/SiC +25
w/Si3N4 +12.5
5) ZnZ04 <1%
w/SiC +2
w/Si3N4 +0.5
6) Li0.Mo03 +3
w/SiC +22
w/Si3N4 +12
MS
Percent Coni
Gases
Formed
to SiOj
None
CO2
12
NO
11
None
—
CO2
12
NO
11
None
--
CO2
60
NO
90
None
--
COj
100
NO
90
None
—
COj
8
NO
4
None
COj
90
NO
90
*'rhe mass changes that occur with SiC and .S'i3N4 represent
the actual increase in mass percent during the oxidation
of the carbide or nitride.
The mass gain in each of these situations is the result of the
oxidation of silicon carbide or silicon nitride to silicon dioxide. Since
significant conversion of the carbide and nitride occurs only with the
lead tungstate, lead cations may catalyze the conversion.
Mixtures of lead phosphate and the carbide or the nitride also
produce large conversions to the oxide. Lead cations may play an
important catalytic role in these reactions as well.
79-6
B.
OXYTHIOMOLYBDATES AND OXYTHIOTUNGSTATES
Table II summarizes the thermal and mass spectral data for the
oxythiomolybdates and oxythiotungstates studied. The mass loss for each
of the compounds is consistent with the amount needed to remove all of the
sulfur from the anions in all of the substances except antimony oxytrithi-
otungstate. In that substance more than 75% of the sulfur is removed so
at least some of the product could be antimony tungstate. Each of the
compounds showed significant interaction with the silicon carbide and the
silicon nitride as evidenced by the oxidation of the ceramics to silicon
dioxide.
Table II Oxythiomolybdates and Oxythiotungstates —
Thermal and Mass Spectral Data
TGA
MS
Mass
Percent
Gases
Percent Conversion
Change*
Formed
to SiO;
CSjMoOS,
-13
SOj
—
w/SiC
+14
CO2
56
w/Si3N4
+5
NO
36
CS2WOS3
-8
SO2
—
w/SiC
17
C02
70
w/Si3N4
8
NO
60
Sbj(WOS3)3
-11
SO2
—
w/SiC
6
CO2
24
w/Si3N4
2.5
NO
17
Zn2MoOS3
-40
SO2
—
w/SiC
14
CO2
56
w/Si3N4
11.5
NO
85
Zn2Mo02S2
-22
SO2
—
w/SiC
8
COj
32
w/Si3N4
5
NO
36
*The changes that occur with the SiC and Si3N4 represent
increases associated with the oxidation of the carbide
and nitride and occur after the oxythiomolybdates and
oxythiotungstates have lost mass.
V.
RECOMMENDATIONS
Analytical studies should be conducted to determine the composition
of the reaction products generated when the oxythiomolybdates and
oxythiotungstates are heated.
Scanning electron microscopy should be used to determine the
crystalline nature of the products.
Infrared spectroscopy should be used assist in the identification of
the reaction products.
Salts containing lead, cesium, zinc, cadmium, antimony, or magnesium
cations could be used in conjunction with the substances examined in this
study to determine whether any of them functions as a catalyst for the
conversion of silicon carbide and silicon nitride to silicon dioxide.
79-8
1990 USAF-UES SUMMER FACULTY RESEARCH PROGRAM/
GRADUATE STUDENT RESEARCH PROGRAM
Sponsored by the
AIR FORCE OFFICE OF SCIENTIFIC RESEARCH
Cojiducted by the
Universal Energy Systems, Inc.,
FINAL REPORT
EFFECT OF EVAPORATION ON THE DRIVING CAPILLARY
PRESSURE IN CAPILLARY PUMPED
AEROSPACE THERMAL MANAGEMENT SYSTEMS
Prepared by:
Academic Rank:
Department and
University:
Research Location:
USAF Researcher:
Contract No:
Kevin P. Hallinan, Ph.D. / David Welter
Assistant Professor / Graduate Student
Mechanical and Aerospace Engineering
University of Dayton
Aero Propulsion Laboratory
Wright- Patterson AFB
Wright- Patterson AFB, OH 45433
Michael Morgan
F49G2U-85-C-UU13
EFFECT OF EVAPORATION ON THE DRIVING CAPILLARY PRESSURE IN
CAPILLARY PUMPED AEROSPACE THERMAL MANAGEMENT SYSTEMS
by
Kevin P. H alii nan
David Welter
ABSTRACT
Research has been conducted to determine the effect of evaporation on the driving
capillary potential in capillary pumped heat transport devices usc'd in aerospace' thermal
management. These devices primarily include heat pipes and capillary pumped loops.
Current design criterion for such devices rely upon what has been termed a maximum
capillary potential to evaluate the maximum heat transport limitations. Preliminary ana¬
lytical results, based upon an idealized model of pores within a heat pipe evaporator wick
indicate that in high powered heat pipes and capillary pumped loops where evaporator heat
fluxes are approaching than 100 VV/cm^ that dynamic forces owing to evaporation from the
liquid-vapor interfaces within the heat pipe evaporator wick can noticeably increase the
driving capillary potential relative to static conditions (if boiling is not occurring). For
apparent contact angles of less than 10° and a.ssumed isothermal inicrfacinl conditions,
normal viscous forces were shown to affect the capillary pressure for capillary numbers
greater than 10“®. This conclusion is particularly true if the working li()uid in these de¬
vices nearly perfectly wets the wick .structure in the vicinity of the licpiid- vapor itit('i faces.
As a verification of the analytical efforts, an experimental facility has been constructed
to actually measure the influence of evaporation on the capillary pK'ssure existing at a
curved, liquid-vapor interface.
80-2
Acknowledgements
I wish to thank the Aero-Propulsion Laboratory and the Air Force Office of Scientific
Research for sponsorship of this research. Universal Energy Systems must be mentioned
for their concern and help to me in all administrative and directional aspects of this
program
My experience this summer, and that of my graduate student, David Welter, was edu¬
cational. Working in laboratories as well equiped as the Thermal Systems group within the
Aero-Propulsion laboratory was exciting. In particular I would like to thank Michael Mor¬
gan, Won Chang, Brian Hagar, John Leonard, Mike Ryan, and Don Reinrnuller for their
help in acquiring and then constructing the experimental facility to measure the dynamic
capillary pressure.
EVAPORATOR
1 Introduction
Heat pipes and capillary pumped loops are perhaps the most signilicaut heat transport
device invented to date. Their applications in ail facets of society are potentially vast. In
the aerospace field, their most attractive use has been related to the rooiing of satellites
where their extremely high heat transport capability is sullicient to handle evaporator heat
fluxes of up to 100 kW/cm^ offers space and weight savings relative to other Types of heat
exchange devices. An excellent review of the application of heat pipes in the aerospace field
is offered by Chang and llagerflj. Another important advantage of heat pipes is that they
do not require any external power sources. They operate i)assively relying upon capillarity
to transport the liquid from the condenser to the evai)orator through the wick. The vapor
returns to the condenser by a pressure gradient arising from a typically small temperature
difference between the evaporator and condenser sections of heat pipes. Figure 1 helps
to illustrate the simplicity inherent in th^operation of heat pipes in their most elemental
configuration.
Despite the apparent simplicity of the heat pipe operation a number of very funda-
804
mental issues associated with the thermal/fluid behavior witiiin these devices is relatively
unknown. Further, the phenomenon of capillarity, wiiich is the principle mechanism inher¬
ent in the heat pipe operation is not well understood. Other problematic issues requiring
further research include: porous media flows; solid-liquid-vapor interaction at the inter¬
faces; and nucleate boiling.
The lack of a proper fundamental understanding of the operation of heat pipes and
capillary-pumped devices is demonstrated by the apparent inability of current models and
design criterion to predict the maximum heat transport capabilities of such devices, even in
steady-state conditions. Gottschlich has given evidence that predicted performance of heat
pipe operation in steady-state is often under- and over-predicted by orders of magnitude
[2]. Obviously, the models predicting transient performance are even worse.
A review of the design literature for heat pipes (Dunn and Reay [.'3], Chi [4], and the
B and K Handbook [5j) all utilized a troubling concept: the maximum capillary pressure.
The maximum capillary pressure is used to evaluate when a heat pipe will fail (i.e., the
evaporator wick dries out and the temperature of the pipe increases significantly). For a
given heat input to the evaporator in steady conditions a certain liquid flow rate within
the wick and vapor flow rate within the vapor core will be cxpcctc'd. 'I'Ik' lu'at pip(' is said
to fail when the liquid and vapor pressure drops associated with the flow becomes greater
than the maximum capillary pressure. Mathematically this is represented as:
(f^cap)„,ar ^ Pa
(1)
Above, S/pg represents the gravitational head that must be o\cr(ome. In microgravity
conditions this term is usually negligible. The problem with this convention is that three
major assumptions are associated with the typical determination of the maximum capillary
pressure. First, it assumes that static conditions exist at the liciuid-vapor interfaces within
the evaporator and condenser wicks. Second, it assumes that a flooded condition is present
within the condenser. Third, it assumes that the liquid perfectly wets the wick within
the evaporator. The first and third assumptions aie not generally valid. The second
assumption has yet to be verified. Thus, the basic premise of the maximum capillary
pressure is questionable. Nevertheless, the collection of the,se assumptions leads to the
following expression for the capillary pressure:
( ^Cap ) rnux
2(7
(2)
80-5
where a represents the surface tension of the interface and r is the mean pore size within
the evaporator.
Given the uncertainty in the maximum capillary pressure, the present research aims
to at least answer the appropriateness of tlie first assumption utilized in its development.
Namely, is it valid to neglect the viscous and inertial forces acting upon a curved liciuid-
vapor interface in the determination of the pressure drop across the interface (i.e., the
capillary pressure)?
2 Objectives of the Research Effort
The goal of this study is to determine how the capillary pressure, (p„ — p;), at a curved
interface is effected by viscous forces resulting from the motion of the fluid due to evapora¬
tion. In particular the investigation will concentrate on determining improved measures to
predict the driving capillary forces in such devices both when film evaporation or boiling
are occuring within the porous wick. Specifically, the research objectives were:
• Show that when film evaporation is occuring within the evaporator in capillary
pumped thermal transport devices that inertial, viscous, and surface tension forces
acting on the liquid-vapor interfaces can affect the pressure difference between the
liquid and the vapor at the interface and, thus, the driving capillary pressure. The
conditions characterizing the importance of these forces in actual heat pipe or cap¬
illary pumped loop operation were to be determined.
• Show theoretically that heat pipe operation can be dramatically improved by in¬
suring that a nucleate boiling regime is stably maintained within the evaporator. A
feasible mechanism is proposed which can explain why the maximum opcarating con¬
ditions of heat pipe devices are affected by the start-up healing |)rolocol. Generally,
higher power transmission is obtained when the heating is increased gradually. This
gradual start-up should allow for stable non-boiling operation and a higher degree
of superheat within the evaporator. Eventually, when boiling does occur, smaller
bubbles will be stably generated, delivering a higher capillary pressure (pressure dif¬
ference between the liquid surrounding tlie bubbles and the vapor in the bubbles).
In gravitational fields buoyant forces are responsible for removal of the bubbles from
the wick. In microgravity applications surface tension gradients acting along the
80-6
bubbles are postulated to be responsible for the bubble migration through the wick
to the vapor core. An investigation of this mechanism will be addressed.
The objectives of the summer research were too ambitious. The work performed during
the summer primarily addresed the first goal, related to determing the dependence of the
capillary pressure on fluid dynamics. The second task described above will be the subject
of an additional AFOSR proposal.
3 Analytical Solution of the Dynamic Capillary Pres¬
sure
This phase of the research concentrated upon the development of a strictly analytical
solution to demonstrate analytically how the capillary pressure changes at curved liquid-
vapor interfaces when film evaporation is occurring. The probability that dynamic forces
can be important at an interface stems from the generally accepted equation representing
the balance of forces at a curved interface normal to the interface. In non-dimensional
terms, this interfacial normal stress balance is given by (See Prosperelti [6])
Pv - Pi + Ca* it ’ {ti - Ar„) • it = A', (3)
where Ca is the capillary number (defined below), r is the dimensionless shear stress
tensor, A is the ratio of the vapor to liquid kinematic viscosity, ft is the vector normal
to the Interface pointing to the vapor side, P is the dimensionless pressure, and f\ is the
dimensionless interfacial curvature. In the typical definition of the maximum capillary
pressure the viscous terms above are neglected. This can only be true if the capillary
number is small.
Regular perturbation theory is employed to determine dynamic contributions. With
this technique velocities, stream functions, and pressures (within the liquid and the vapor)
are all expanded in terms of the capillary number, Ca, defined as
Ca =
(7
(4)
Above refers to the kinematic viscosity and U refers to the average velocity normal to
the interface. The subscript / refers to the liciuid. The significance of the capillary number
is that it has been shown to adequately icprcsent the ratio of viscous to surface tension
forces acting upon an interface thus it represents a reasonable parameter with whicli the
velocities and pressures can be expanded. For example:
Pv,i — {Pv,i)o + Ca{Pvj)i H -
^v,i = (^v,i)o + Oa(u^j)i -{ -
Actual values of the capillary numbers present near evaporating interfaces in real heat pipes
range from 10"® in water heat pipes to as high as 10"® in lithium heat pipes. Apparently
in cill C2ises the capillary number is small. However, such values do not necessarily imply
that viscous forces are unimportant relative to surface tension forces. For example in free
convection, buoyant forces are only important relative to viscous forces when the Grashof
number, defined as a ratio of buoyant to viscous forces, is on the order of lO’^. Thus, the
capillary number does not have to be on the order of 1 to represent a balance of viscous
to surface tension forces.
The physical model of the evaporating interface focuses on just one pore within the
evaporator wick, and particularly the contact line region. A wedge type model of this
region, as shown in Figure 2 is utilized similar to that employed by Huh and Mason (7) and
Cox [8] in describing the unrelated problem of an advancing (non-evaporating) liquid-vapor
interface within a capillary tube. The liquid-vapor interface is assumed to be isothermal
(a realistic assumption considering the typically small pore sizes on the order of 10"® m,
making it likely that the liquid and pore structure temperature will locally be uniform). As
a consequence, the evaporation rates are uniform over a majority of the interface. Only near
the wall where Derjaguin[9j and WaynerflOj have shown that solid-liquid intermolecular
forces for wetting fluids locally reduce the liquid pressure near the contact line and thus
choke off the evaporation rate. Thus, at the wall the evaporation rate is zero. However,
the lengths, s, at which these long-range intermolecular forces are important is only on
the order of 50 nm (Potash and VVayner (IlJ) and thus this region is small relative to the
entirety of the pore. It is argued by Kafka and Dussan (12), that an imprecise treatment of
this near contact line region affects only slightly the flow and pre.ssure fields outside of this
region. Thus, the model employed utilized a uniform evaporation rate over most of the
interface. At a distance s from the apparent contact line the evai)oration rate is assumed
to experience a linear decline from tlie uniform rate to zero at the apijareiit contact lino.
(Note that there is a wealth of literature relevant to the exact definition of the apparent'
contact line.)
80-8 '
uniform evaporation
rate
Figure 2: Wedge model used for the analysis.
Steady-state conditions are considered. Plane-polar coordinates are used, implying that
the circumferential curvature of the pores is small. Low Reynolds and capillary number
flows typical of heat pipes are considered. A review of a significant number of papers
describing heat pipe performance has revealed that capillary numbers up to about 10“ 5
are observed in liquid metal heat pipes and up to 10“7 in low Icmperatiire heat j)ipes. Pore
Reynolds numbers of below 10“5 are universally observed. Because of these extremely
low pore Reynolds numbers as.sociatcd with the cvai)or<ilor li(iui(l (low i.h(i lubrication
approximation is used to describe the flow near the contact line within a representative
pore in the evaporator wick.
With these assumptions the momentum equation in each phase is written as :
- Vpj.„ + p/.v = 0 (5)
The momentum equation is made dimensionless by scaling the lengths with /?, the pore
radius, and the velocities with U, the average liquid velocity at the interface, and dividing
by the surface tension, a. In dimensionless form the liquid and vapor momentum equations
80-9
are: (neglecting body forces which could easily be accounted for).
(6)
(7)
(8)
The asterisk implies dimensionless quantities, A represents the ratio fivlf^h and Ca is the
capillary number. The velocity and pressure fields are expanded in terms of the capillary
number for small values of the capillary number. The asterisk is dropped but all quantities
are to be considered dimensionless.
u = uo + Caui + • • • (9)
P = Po + CaP, +••• (10)
Notice that in the dimensionless form of the momentum equation the viscous terms are
multiplied by the capillary number. Thus, at 0{Ca^°) the only term present is the pressure
term (and the body force term). As expected if tliere is no flow the pressure field is governed
by hydrostatic conditions. Since our interest is in determining the first-order correction to
the capillary pressure, we need only compute the zeroth order velocity fields.
Near the contact line the velocity fields within the liciuid and vapor phases are subject
to the following boundary conditions:
• At wall:
- VP,* = 0
ACuV^T; - vp; = 0
- no-slip and impermeability . (11)
• At interface:
— continuity of evaporative (lux (12)
— continuity of liquid and vapor tangential velocities (13)
— continuity of tangential shear stress (14)
(15)
Additionally the conservation of normal niomeintum at the inteiface yields an expression
relating the liquid and vapor pressures at the interface.
80-10'
The two-dimensionality of the problem allows for the representation of the velocity
fields in terms of stream functions.
d'Hi
and
i^ih = -
dr
(16)
The stream-functions can be likewise expanded in terms of the capillary number so that
the zeroth order velocities are related to the zeroth order stream functions as shown:
(«/,o)r = --
and = --
(17)
r -
Taking the curl of the liquid and vapor momentum equation and representing the velocities
in terms of the stream functions for both phases, yields at zeroth order the biharmonic
equation:
(18)
The no-slip boundary conditions for u/g and at the wall gives:
dH
= 0
= 0
d<j>
di^
= 0 on <^ = 0,
= 0 on (f>= TT.
(19)
(20)
(21)
The evaporative flux boundary condition, represented according to Figure 2, is given by:
dr dr
= - - + //(r-e)(l--)
e j
where H is the Heaviside function. Continuity of tnng('ntini velocity re(|uircs that:
a'k/o a'P,
VO
at (t> = 9o
d4> d(i>
The tangential shear stress continuity at the interface is represented by:
r^ d(f>'^ dr"^ r dr
dr^
-f -
dr )
The balance of the normal stresses on the interface is given dimensionally by:
Pv - Pi + » • (n - Ty) • ?7 = crK
(22)
(23)
(24)
(25)
The curvature is defined by K = dOfds, where 9 is the angle of tangency to the interface
and s is the arc length. In the wedge approximation for the meniscus, s r, and thus,
.. d9
The normal stress equation then can be represented non-dimensionally by:
P^-Pi + 2Ca
4. 1^1 - (-1^ 4
rd^idr H d(f> rd(i)dr d<i>
dl
dr
(27)
Since the angle of tangency, 6 = Oo + Ca0i{r)-\ - , is assumed planar at 0(Ca'^°), only at
©(Ca"^*) is any curvature considered and is a result of the viscous forces. Notice again that
at zeroth order only the pressure and surface tension are of consequence. Thus, neglecting
body forces, at zeroth order, away from the thin film region the i)re.ssure drop across the
interface is controlled by curvature. At first order the normal stress equation is given by:
“ ^<1 + 2
r d(f)dr d<f> r d(i>dr d<j)
dr
(28)
A solution to the biharmonic equation for both phases satisfying the boundary equa¬
tions at zeroth order is given by:
^/.vo = f’l(Avo^ + 5/.tw)cos?^ + (C'j,to^^ + A.vo)sin9^] (29)
-f (G/.vo-^ + ^o)sin^] (30)
(31)
The constants through are obtained from the imposition of the boundary con¬
ditions in both the liquid and the vapor.
With the stream functions determined tlie velocity fields were computed. The pressure
fields were then determined from the momentum equations up to a constant. A solution
for the liquid and vapor pressure fields of tlie form shown was used.
Pl.» = + Kl: (^2)
The functions and were found to be:
= 2Ai,^cos{<j>) + 2C(,^Qs\n(i> (33)
ii.v, = £i,voi-cos(^ - ^siiKf)) -
+ G u.s,{<f>cos<j> - sim^) -f Hi^i^cos<f> (34)
The prime interest in the study lies in the calculation of the pressure jump across the
interface (P„ - Pi). Thus, it is only necessary to solve for the dilference in the constants,
Kv — A’/, which can be defined to be auotlier constant, A;,,. I'lie constant A/„ can be
80-12
Flow
liquid
1 Ki
Wall
O “ Face or side of control volume
Figure 3: Control volume straddling contact line used for force balance.
determined by performing a simple force balance on a control volume straddling the contact
line. The balance of forces in the flow direction yields in dimensional terms the following
expression.
fH fH (Q Qiii r-L Oil
/ pi^l\x=Ldy+ Pilx^idy - pi—\y=odx - Pv-^U=odx
Jo Jo Jl uy Jo ay '
/’** dui dux,
+ ^^~Jl ~ PvUvVvly^ndx
- [ Pv\y=Hdy - I Pv^illxc^idy = 0 (35)
Jo Jo
The inertial terms on faces A, C, and D for small Reynolds number flows are all
negligible relative to the pressure terms and the viscous terms acting at the wall and on
face C. The following simplified equation results.
~ II ~ lo Jl Jj>
80-13
+ / Pilx^idy - f Pv\x=-Ldy = 0
Jo Jo
(36)
Expanding the velocities and pressures in terms of the capillary number as before =
(7au/,„j H — , etc.) and substituting into the force balance equation yields expressions of the
force balance at different orders in the capillary numl)er. At order zero, static conditions
(assumed known) exist. Dropping the bar depicting dimensionless parameters the zeroth
order term is:
I Pifi\x=Ldy - I Pvfi\x=-Ldy = 0 (37)
JO JO
At first order the following equation results:
dy
du^,o I
dy
+ / Pi,i{x = L)dy - f Pv.iU=-Lrfj/ = 0
Jo Jo
(38)
Thus, the first order correction to the capillary pressure depends again on the zeroth order
correction for the velocity field which is known. Additionally the zeroth order corrections
for the vapor and liquid pressures are known up to the constants A'j and K^,. The constant
Kiv is chosen so that the equality is satisfied. This comes from the fact that the integral
given by:
I” PvU-L - P,\T=Ldy (39)
Jo
can be written in terms of a known part plus an unknown contribution due to the constants.
Simply, this is as shown.
rH rll
/ (Known part of (P„ - P; )](///+ / h’lvdy (40)
JQ Jo
Thus Kiv is determined by requiring Equation 38 to be satisfied.
In summary the velocity, and pressure fields are computed at zeroth and first order
in capillary number from the Navier- Stokes equations, subject to standard boundary
conditions. The lubrication or low Reynolds number approximation of the Navier-Stokes
equations is employed. Exact solutions for the velocity and pressure fields were obtained.
Results
A sampling of the results from this study are included in this repoit. Figure 4 shows
the streamlines (dimensionless) for a case where the dynamic contact angle is 10 degrees
and the kinematic viscosity ratio is O.OI. The licpiid region is to the light of the interface
80-'l4
X
Figure 4: Representative flow streamlines near contact line at a dynamic contact angle of
10 degrees.
line. The vapor region is to the left. It is interesting to observe the curvature of the liquid
streamlines toward the solid wall as the interface is approached, owing to the requirement of
a uniform evaporation rate. This trend is expected. Upstream of the interface a parabolic
velocity profile might be expected due to the representation of the pore as a wall, and
because of the shear stresses at the wall. Due to the uniformity of the evaporation rate,
the liquid velocity profile must undergo a transition from parabolic to near uniform, and
thus, liquid must flow toward the wall. At the interface the liquid transitions to vapor.
Continuity of tangential velocity between the liquid and vapor and the requirement of a
continuous normal evaporative flux at the interface necessitates that the vapor upon leaving
the interface must initially flow toward the wall. Downstream, however, the presence of
the wall decelerates the flow causing the vapor streamlines to diverge from the wall.
What is more interesting is how thgse streamline trends affect the capillary pressure.
Consider that in static conditions there is no flow. The pressure drop (from vapor to liquid)
is determined from a static pressure balance yielding the expression for the maximum
capillary pressure described earlier. Now, in order to have a uniform evaporation rate the
80-15
liquid has to move toward the interface. Thus the liquid pressure at the interface has to
be reduced. Additionally, in order for the vapor to move away from the interface, the
vapor pressure at the interface must increase above its static value. Thus, the dilfercncc
between the vapor and liquid pressure across the interface must increase. This trend is
shown in Figure 5 where the first order correction in the dimensionless capillary mmibcr,
(Pca|))ij is presented as a function of the distance from the apparent contact line. A
viscosity ratio of 0.01 is common to all four curves, representing data for 2, 5, 10, and 20
degree apparent contact angles. Two trends are observed. First, in all cases the capillary
pressure increase at first order declines with increasing distance from the contact line.
More importantly, the capillary pressure correction increases rapidly as the contact angle
decreases. For a 2 degree contact angle, values for (Pcap)i ranged from 2x10^ at li = 0.01
to a near constant value of over 10“* at large R, These numbers are significant relative
to the zeroth order or static dimensionless capillary pressure which is on the order of
1. Capillary numbers in heat pipes of up to 10~® have been observed. Since the total
capillary pressure existing at an interface is the contribution of the static plus the first
order: Reap = (Pcap)o + Ca(Pcap)i, then if (Pcap)t is in the range reported for the smaller
contact angles, it can be concluded that the fluid dynamics of an evaporating interface
can in fact increase the capillary pressure existing at the interface. Tlius, in high-powered
heat pipes, it is argued that dynamic effects might be incorporated into the definition of
the maximum capillary pressure. However, it is recognized that the exact correlation of
the model used in this study to real heat pipes is sketchy. The results merely show that
for apparent contact angles less than 10" that dynamic ellects can be significant when the
capillary number is on the order of 10“®. The experiments described below will help draw
a better correlation.
4 Facility to measure the capillary pressure
A facility has been constructed and assembled to afford the opportunity to actually measure
the capillary pressure exsiting across a curved liquid-vapor interface. A schematic of the
facility is shown in Figure 6. A vertical capillary tube is mounted within an evacuated
pressure vessel. A C02 laser beam, directed upon the interface, is to be used to affect
evaporation. The power of the C02 beam is variable. A large resc'i voir is to supply liciuid
to the capillary tube. Another capillary tube will be used to connect the reservoir and the
80-16'
Contact Angle
R
Figure 5: First order correction to the dimensionless capillary pressure dependence upon
the distance from the interface.
evaporator, across which, the change in pressure will be measured. A cathetometer is to be
used to measure the difference in levels between the reservoir and the meniscus within the
capillary tube. The vapor pressure above the reservoir and within the evaporator are to
be measured. Thus, in stable evaporation, the pressure juini) across the meniscus can be
compared to the pressure losses between the reservoir and the evaporator. The capillary
pressure can then be inferred from the measurements. The pressure drop associated with
the flow will be determined from the knowledge of the flow rate and the assumption of
fully-developed conditions within the interface. Mathematically, this is written as:
Pcap — -Pc ~ -Fr + AP horizontal tube vertical tube (41)
In this equation Pcap is the capillary pressure existing across the evaporating interface, P<. is
the vapor pressure within the chamber housing the vertical capillary tube, P^ is the vapor
pressure above the reservoir supplying liquid to the capillary tube, AP
and AP vertical tube pressure drops, and pig Ah is the head difference between
the reservoir and the meniscus.
These experiments will i)e performed for a vaiiely of evapoiafion rales, including,
80-17
U) irr>i
see j low
LASCH OCAMS
CAtMl l AIIY Mini U)
ALLOW f OK I low hail
MLASUm III HI
Figure 6: Schematic of experimental facility to study elfccts of Iluid dynamics on capillary
pressure.
situations where boiling is occuring. A microscope will be used to observe the evaporating
meniscus to additionally observe- how the shape of the interface changes with increasing
evaporation rates. It is noted that these experiments are similar to those performed by
Wayner [10]. However, a significant difference exists. VVayner did not measure the dynamic
capillary pressure inditectly as is proposed here. Instead, he ijhotographed the meniscus
at varying evaporation rates. From these photographs he inferred the dynamic curvature,
K^vap- The dynamic capillary pressure was then related to this dynamic curvature using
Laplace’s equation (which is valid only for static conditions). This reduction of the data
was incorrect, as shown by Equation (3). The normal viscous stresses ultimately are
responsible for the change in the capillary pressure.
The facility is now complete and ready for testing. The mini-grant is a necessity to
allow for the opportunity to complete the experiments.
5 Recommendations
The work described in this report represents only the beginning of a total research effort
deemed necessary to gain a clear understanding of the driving rapillaiy potential in heat
pipes. A methodical approach is suggested beginning with simple idealized problems and
gradually approaching physical geometries more realistic of actual heat pipe wicks. First,
support is essential for the completion of the experiments with the existing facility, looking
at evaporation from a single capillary tube. Next, the vertical capillary tube should be
replaced by a larger diameter tube. Real wick materials will be inscitcil into the tube. The
C'02 beam will be expanded to fall nearly uniformly over the top of the porous material.
The dynamic capillary pressure will once again be inferred from the measurements. These
results would definitely constitute a significant contribution to heat pipe designers, in that
for real heat pipes, for evaporation rates typical of heat pipes, the dynamic elfects on the
capillary pressure will be known.
Of course, additional work is also needed to investigate the condensing problem at
a curved interface to assess whether the Hooded assumption Upically used in heat pipe
condensers is valid. A similar project can be performed to deteimine the influence of
condensation of the capillary pressure existing across the li(iuid-\apor interfaces within
the heat pipe condenser wick.
.A number of additional studies aie necessary to fully undei stand the phenomenon of
30-19
capillarity within heat pipe wicks. These and other questions must be answered. How
is the driving capillary pressure effected by the occurrence of nucleate boiling within the
evaporator wick? If bubbles are generated within the evaporator wick, is it possible for
these bubbles to migrate from the wick? What is the transient behavior of the driving cap¬
illary pressure during start-up conditions? These questions must eventually be addressed
to overcome the problems currently inherent in the use of heat pipes and capillary pumped
loops.
References
[1] Chang, W. S. and Hagar, B. G., ASD90 1490 Technical Report, 1990.
[2] Gottschlich, J. M., 1989 SAE Aerospace Tech. Conf., Anaheim, CA.
[3] Dunn, P. and Reay, D. A., 1982, Heat Pipes, 3 ed., Pergamon Press, New York,
New York.
[4] Chi, S. W., Heat Pipe Theory and Practice, MacGraw-Hill, 1967.
[5] Brennan, P. J., B and K Heat Pipe Design Handbook, B and K Engineering,
Inc., Towson, MD, June 1979.
[6] Prosperetti, A., 1979, Meccanica, p. 34.
[7] Huh, C. and Mason, S. G., 1977, J. Fluid Mech., 81, p. 401.
[8] Cox, R., r., 1986, J. Fluid Mech. ,168, p. 169.
[9] Derjaguin, B.V., 1955, Colloid J. USSR, 17, p. 207.
[10] Wayner, P. C., Jr., 1978, J. Heat Transfer, 100, p. 100.
[11] Potash, M., Jr. and Wayner, P. C., Jr., Lit. J. Heat Mass Transfer, 15, p. 1851.
[12] Kaflca, F. Y. and Dussan, V., E. B., 1979, J. Fluid Mech., 95, p. 539.
1990 USAF-UES SUMMER FACULTY RESEARCH PROGRAM/
GRADUATE STUDENT RESEARCH PROGRAM
Sponsored by the
AIR FORCE OFFICE OF SCIENTIFIC RESEARCH
Conducted by the
Universal Energy Systems, Inc.
Fm.BEPO.BT
Investigation of the Combustion Characteristics of a Confined Coannular
Jet with a Sudden Expansion
Prepared by:
Paul O. Hedman, Ph.D., P.E.
Academic Rank:
Professor
Department and
Chemical Engineering
University:
Brigham Young University
Provo, Utah 64602
Research Location:
Aero Propulsion Laboratory
Wright-Patterson AFB, OH 45433
USAF Researcher:
W. M. Roquemore, Ph.D.
Date:
20 July 1990
Contract No:
F49620-88-C-0053
Investigation of the Combustion Characteristics of a Confined Coannular Jet
with a Sudden Expansion
by
Paul O. Hedman, Ph.D., P.E.
Chemical Engineering Department
Brigham Young University
Provo, Utah 84602
ABSTRACT
This report contains a brief summary of the work done to investigate the
operation al characteristics of a burner that was designed to "specifically reproduce
recirculation patterns and LBO processes that occur in a real gas turbine combustor." Measurements of
lean blowout limit were conducted. The lean blowout limit was found to closely correspond to the lean
flamability limit found in a well stirred reactor. LDA measurements of the gas velocity were made to
determine the pattern of the complex flows In the combustor, and to identify regions of recirculation flame
flow patterns. Measurements were made using spontaneous OH uv emission to characterize the fuel
equivalence ratios where the flame was attached to the bluff step just outside the air jet and where a fully
lifted flame occurred. At very fuel rich fuel equivalence ratios, the flame was attached to a greater or lesser
degree. As the fuel equivalence ratio approached about 1 .06, there was a point of demarcation where the
flame became full lifted. Two-dimensional images of the flame and flow structure were taken with laser
sheet lighting and a CCD camera. Those images have shown the complex nature of the shear layers
between the outer recirculation zone, the air jet, and the fuel jet. The eddies in these shear layers are of
about the same scale as the annular space between the edge of the air jet and the fuel tube.
81-2
ACKNOWLEDGEMENTS
I wish to thank the Aero Propulsion Laboratory, Wright Patterson Air Force Base, and the Air Force
Office of Scientific Research for sponsorship of this research. The timely help of Mr. Rodney C. Darrah
and other individuals at Universal Energy Systems was greatly appreciated. Special recognition of the
help of Mr. Milton H. Danishik (UES) is acknowledged. Not only did he provide extraordinary help in
processing my bill for services, but he became a good friend as we shared the same office.
The very successful experience gained during this summer research program is in large measure
due to the very excellent support of my sponsor. Dr. W. M. (Mel) Roquemore (WRDC/POSF), and the very
well qualified research staff at Systems Research Laboratory, Inc. (SRL) with whom I had the pleasure of
working this summer. Special thanks are in order for Dr. Larry Goss, Dr. Viroj Vilimpoc, Mr. Benjamin Sarka,
and Mr. Michael Post, of SRL Their willingness to help, to answer questions, and the help that they
provided made this summers effort truly rewarding.
81-3
I.
INTRODUCTION:
This report presents a brief summary of resuits of an investigation to determine the fiow and fiame
characteristics of a burner with a confined, coannuiar jet with a sudden expansion (Pratt and Whitney Task
too Combustor). which has been deveioped to study the phenomenon of iean biowout (LBO) in modern
annuiar aircraft gas turbine combustors. The combustor has been carefuiiy designed (Sturgess, et ai.
1990) to "specificaliy reproduce recircuiafion patterns and LBO processes that occur in a reai gas turbine
combustor."
The combustor consists of coaxiai jets with a 29 mm diameter centrai fuel jet surrounded by a 40
mm diameter annuiar air jet. The jets are located in the center of a 150 mm diameter duct. A sudden
expansion, reamvard facing bluff body, with a step height of 55 mm, is iocated at the exit plane of the
coaxial jets. The combustor test section incorporates flat quartz windows to accommodate laser and other
optical access, but uses a metal shell with metal corner fillets to reduce the vorticity concentration and
eliminate its effect of the bulk flowfield in the combustor. This box-section combustor with corner fillets
allows reasonable optical access, while providing a cross section that approximates a two-dimensional
axisymmetric cross section. The bluff body provides a recirculation region that can stabilize the flame.
When operated in a very fuel rich mode, the flame is very stable and is attached to the bluff body
near the outer edge of the air tube. As the fuel equivalence ratio is reduced, the flame becomes less
stable, and eventually reaches a point where it lifts from the base region, and becomes stabilized on the
outer recirculation zone. Thus, there are two distinct operating modes for the burner, a fully attached
flame, and a lifted flame. As the fuel equivalence ratio is further reduced, the flame becomes very
unstable, and eventually completely blows out of the burner. This lean blowout limit has been found to
correspond to a fuel equivalence ratio of about 0.49, very close to the lean flammability limit of the propane
fuel and air (ca ^ = 0.50, Lewis and von Elbe, 1987). Other studies (Sturgess, et al. 1990; Longwell, et
al., 1953) suggest that the combustor is behaving like a well stirred reactor when it is near its lean blowout
limit.
The objective of the project was to determine the combustion and flow characteristics of the
burner over a range of operating conditions. Specifically, the study was to examine the mechanisms that
contribute to the transition between a flame attached to the bluff body of the sudden expansion, and a
lifted flame. These characteristics relate to a flame blowout modeling study being conducted by other
investigators from Pratt and Whitney (Sturgess, et al. 1990). A secondary purpose of this study has been
to collect data that could be used in validating a computer code that predicts the flame phenomena and
blowout limits (Sturgess, et al. 1990).
The study has developed a better understanding of the overall behavior of the combustor, has
investigated those phenomena that contribute to the transition from a flame that is well attached to a lifted
814
flame that is being stabilized on the outside recirculation zone, and has examined how the variation in fuel
equivalence ratio affects the operation of the burner and the lean blowout.
II. APPROACH:
The approach used in this study was to investigate the flow and flame structure within the burner
over a wide range of fuel equivalence ratios. The characterization involved measurement of the fuel
equivalence ratio at lean blowout, LDA measurements of the gas velocity throughout the combustor with
and without combustion, laser sheet flow visualization of the combustion flows using TiCl4 seed and uv
laser sheet lighting to stimulate OH LIF emission, and spontaneous OH emission measurements in the uv
to determine quantitative measurements of the fuel equivalence ratio where the flame transitions from an
attached flame to a lifted flame.
III. RESULTS:
Lean Blowout Limit Measurements. The lean blowout limit (i.e. the fuel equivalence ratio where
the burner extinguishes) was determined by setting the air flow in the combustor, and then by slowly
reducing the fuel flow rate (gaseous propane) until the reactor would extinguish. The fuel flow at blowout
was somewhat dependent on the time temperature history of the combustor. The reactor would sustain a
flame to somewhat lower fuel (equivalence ratios after the reactor had reached thermal equilibrium.
Consequently, the data reported were taken after sufficient time to allow thermal equilibrium to be
reached. Undoubtedly, this thermal effect has lead to some of the data scatter and uncertainty in the data.
The Pratt and Whitney Task 100 Combustor has been designed to accommodate several different
exhaust extensions and exit restrictions. In the configuration tested in this study, the 1 0 inch exhaust
extension was used with and without a 45% exit restriction.
Lean blowout experimental results were measured over a range of operating conditions and
reactor configurations. Results for two combustors (SRL Building 450 combustor, and UDRI Building 490
combustor) with measurements taken at the two different facilities were compared Both combustors were
tested with the 1 0 inch exhaust extension, but other hardware variations included tests with or without the
45% exhaust restriction, and tests with different arrangements of metal wall plates or quartz windows
installed. Example results are presented in Figure 1 . These results show the effect of the 45% exhaust
restriction on the fuei equivalence at blowout. Quadratic least squares fits of the data are shown. There is
a slight decrease (3,4% at 1500 sipm air to 5.5% at 4000 sipm air, an average effect over all air flow rates of
about 3 9%) in the fuel equivalence ratio at blowout when the combustor is operated with the 45%
exhaust restriction While the trend seems to be quite consistent over the range of air flowrates tested,
the differences are generaily within the experimental band of uncertainty, and the conclusion that there is
a significant effect of the 45% exhaust restriction must be used with caution.
81-5
Air Flowrate, sipm (70 F)
Figure 1 - Effect of Exit Restriction (45% Orifice)
on Lean Biowout Limit
Demarcation of Lifted Flame Region.bv Spontaneous OH uv Emission Measurements.
Photographs of the flame clearly illustrate the different attached and detached flame operational regimes
of the burner. However, visual observation and conventional photographic techniques were unable to
quantify the precise fuel eqtiivalence ratio where the flame transitioned from an attached flame to a
detached, flame. Measurements of OH emission from the flame in the region of the attachment point were
made to determine change in flame structure as a function of fuel equivalence ratio. These
measurements have clearly identified the two operational regions of the flame, and have quantified the
point of demarcation between the regimes of attached and detached flames. FFT analysis of the OH
emission data have also provided an indication of the frequencies of the fluctuations in the flame
associated with the attachment phenomena. LDA measurements of gas velocity in the flame at fuel
equivalence ratios corresponding to a fully detached flame, to the point of demarcation between the
detached flame and the attached flame, and to a well attached flame have been made, and used to
81-6
quantify the characteristics of the major recirculation zones, and the mixing shear layers. The influence of
fuel equivalence ratio on the flame structure has been determined.
The burner operates in either a mode where the flame Is attached to the bluff body just outside of
the air annular jet, or in a lifted flame mode. Visual observations of the flame as fuel equivalence ratio was
changed indicated that there was a point of demarcation in ^ between these two regions where the flame
was always lifted. At fuel equivalence ratios in excess of this value, the flame seemed to be lifted part of
the time, and attached part of the time. Example results from these emission experiments are shown in
Figure 2, In an effort to quantify the exact fuel equivalence ratio wl;ere this occurred, an optical system
was set up that would look at a region of the flame near the attachment point.
Figure 2 - Flame Attachment Data, 1000 sipm (70 F) Air flow,
45% Orifice on Exit (Calibrated Phi)
and measure the spontansous OH emission in the near uv at about 308 nm. The optical system was a line
of sight system that had a very narrow view angle, that was focussed through the flame attachment region,
81-7
but nevertheless, integrated the emission signal completely across the duct. Figure 2 iliustrates cleariy,
that there are two distinct regions of operation, and that there is a definite point of demarcation between
these regions. Simiiar measurements at other air flow rates have shown that this point of demarcation
occurs at a fuel equivalence ratio of about 1 .06 independent of the air flow rate.
LDA Velocity Measurements. A laser Doppler anemometer (LDA) was used to make extensive
measurements of gas velocity in the burner. Measurements were made in coid flow simulations of the
flame, and in flames at two fuel equivalence ratios {(|) = 1 .05 and ^ = 1 .56).
The air flowrate used was 1000 sipm (70 F), and the fuel (propane) flowrates used \A ere 41 sipm (0
C) and 63 sipm (0 C). The cold flow simulations of the flame also used 1000 sipm (70 F), but used either
air, nitrogen, or CO2 in the fuel tube. The cold flow measurements were primarily completed to resolve
operational difficulties in the burner, and in the LDA system.
Higher flowrates were desired in the burner, but heatload limitations of the exhaust hood
prevented the continuous operation needed for LDA measurements to about 70 sipm (0 C) of propane.
Consequently, a set air flowrate of 1000 sipm (70 F) was established for the experiments, and fuel flow
rates were selected at the point of d.^marcation between a lifted flame [fuel flowrate of 41 sipm (0 C)
propane, <[> = 1 .05] and a well attached flame [fuel flowrate of 63 sipm (0 C) propane, ([) = 1 .56). Other fuel
flow rates were used for some of the imaging measurements, notably fuel flowrate = 30 sipm (0 C) =
0.78) and 52 sipm (0 C) (<!»=. 1 ,31 ) which correspond to a well lifted flame, and a moderately attached flame
respectively. It would be desirous to have complete sets of data at all four fuel equivalence ratios
considered, but the amour of data that could be taken during the short time of the study was limited to
the measurements at ([» = 1.05 and <!> = 1.56.
LDA measurements were taken in two different zones in the reactor. Due to the limitations in
optical access, and the orientation of the laser beams for the two component LDA, axial and radial velocity
components could be obtained from about -31 mm to +31 mm in the radial X direction. In the other
coordinate direction, axial and tangential velocity data could be obtained from about -70 mm to +70 mm in
the Y coordinate radial direction. Normally, radial velocity profiles were only taken on one side of the duct,
from the centerline out. Complete profiles were usually taken for the data near tiir iet exit to determine
the boundary condition to the reactor, and to determine tht symmetry of the inlet flow conditions. Radial
velocity profiles in either quadrant were take from abo”*. 6 mm ox!di location (closest position possible to
the bluff body sudden step) to 300 mm axial location. Most of the interesting flair, 0 structure existed
between these two axial locations. The axial and radial velocity components are the much preferred
components to evaluate the flows in the axisymmetiic flowfieid.
81-8
The velocity measurements permit the flow structure in the burner to be quantified. One example
set of data is shown in Figure 3. This figure presents data from the near fieid of the burner (Z = 6 mm) out
to a 200 mm axial location. In the near field of the burner, the high velocity associated with the fiow
through the air annulus Is clearly evident. A more complete profile taken at 1 mm steps in the other
coordinate direction shows that the velocity across the air duct is relativeiy flat at about 34 m/s. from about
14.5 mm to 20 mm radius. The course grid of this data set shows only the general details of this region.
The velocity profiles in the fuel tube clearly show that there is a recirculation bubble in the fuel tube that
extends from about 10 to about 32 mm. The exact size of this recirculation zone depends on the fuel flow
Duct Radius (R, mm)
Figure 3 - LDA Axial Velocity Measurements in the
Pratt and Whitney Task 100 Combustor
1000 sipm Air, 41 sipm Propane, Phi = 1.05
rate The change in velocity profiies with increased axial distance is cleariy evident in Figure 3. It can be
seen that the flow in the channel is flattening, and is expected to approach a typical turbulent flow profile
by the reactor exit Preliminary analysis of the velocity data obtained shows that there is a very complex
recirculation zone associated with the bluff body sudden step.
81-9
Laser Sheet Lighting Imaoina Measurements. Two different schemes were used to examine the
flow and combustion detaiis of the flame, TiCl4 seeding and LiF emission of OH. In each scheme, a sheet
of planar laser light was used to capture an instantaneous (10 ns) two dimensional image of the flame and
flow structure. In each case, the plane of the laser light bisected the test section of the reactor, passing
through the centeriine of the reactor. The iaser iight came through the window on the X coordinate. The
image of the iiiuminated fieid was taken by a CCD camera iocated on the Y coordinate. Two different CCD
cameras were used, a 1024x1024 array unintensified camera for the more intense TiCI4 seeded
experiments, and a 384x576 intensified CCD camera for the iess intense Images of the OH LIF. The CCD
cameras were able to capture the full width of the optical window. The camera used with the TiCI4 seeded
experiments was able to capture a vertical image in the quartz window about 1 45 mm in height. The
camera used in the OH LIF imaging was able to capture an image about 79 mm in height. In order to get a
more complete record of the OH images, the burner was located a four different axial locations, 0 mm, 75
mm, 1 50 mm, and 225 mm.
In the first technique, a sheet of frequency doubted Nd/Yag laser light (532nm) was focused
through the test chamber. Mie scattering of this laser light from particles of Ti02 was used to obtain
instantaneous images of the flame and flow patterns. An example of one of these images is shown in
Figure 4. The Images were recorded on a 1024x1024 array CCD camera. The Ti02 seed was . btalned
from the near instantaneous reaction of TiCl4 vapor and moisture to form the Ti02 particles, and HCI
vapor. The TiCl4 vapor was seeded with either the dry air flow or the propane fuel. Moisture was added
with the opposite flowing stream to provide ample H20 for the chemical reaction. Moisture generated in
the combustion process also participated in the chemical reaction. This figure shows the results of
seeding the TiCl4 vapor into the air stream, and adding moisture to the fuel tube. The flow structure
associated with the flow through the fuel tube are evident. The eddies in the shear layer between the fuel
jet and the annular air jet are quite evident. The turbulent structure downstream is also clearly evident.
In the second imaging technique, the laser light from a tunable dye laser was used to form a sheet
of laser light. The frequency used was about 283 nm. This frequency was used to pump a fluorescence
transition of OH which resulted in a LIF signal at about 308 nm. Two dimensional images of the flame zone
shown by the OH LIF were obtained with a 384x576 intensified CCD camera. These images have been
used to examine the iocation and shape of the flame zones in the burner. An example of one of these
images in shown in Figure 5. This image is located in the near region of the fuel and air jets, and clearly
shows the combustion zones associated with the flame. The convoluted shape of the flame front clearly
shows the complex structure of the flame and the associated eddies.
Figure 5 - Example of Laser Sheet image of OH LIF
1000 sipm Air, 63 sipm Propane
IV. DISCUSSION OF RESULTS (CONCLUSIONS^:
Lean Blowout Limit Measurements. The following preliminary conclusions have been drawn from
the results of the lean blowout limit measurements.
1 . In general, the fuel equivalence ratio at lean blowout is independent of air flowrate. Some of
the data, particularly the UDRI data suggest a slight decline in (j) with increasing air flowrate, while other sets
of data, particularly the SRL data shows a slight increase in ([» with increasing air flowrate. The noted slight
increase or decrease are well within the data scatter and measurement uncertainty, and are probably not
meaningful.
2. The fuel flow meters are only readable to 1 sipm of propane at the SRL facility, and 1 sipm of air
(times 0.36 to convert to propane) at the UDRI facility. This translates to about 1 part in 20 to about 1 part
in 80 over the range of propane flow rates being used, or about 5% at 1000 sipm air flowrate down to
about 1.3% at 4000 sipm air flowrate. Trying to determine low percentage trends within these error bands
is risky.
3. There is a slight effect observed in the performance of the burner with and without the 45%
exhaust restriction installed. The 45% exhaust restriction lowers the fuel equivalence ratio at lean burnout
by an average of about 3.9% over the range of conditions tested. The effect seems consistent, even
though it is of the same order as the scatter and experimental uncertainty in the data.
4. There was no significant effect observed on the effect of quartz windows or metal plates
installed in the reactor. Limited data taken on the SRL facility showed a small effect, but a more exhaustive
set of data taken on the UDRI facility over a much wider range of test conditions showed the the effect was
insignificant.
5. There was a slight difference in the performance of the SRL and UDRI burners, that could not
be explained. In general, the fuel equivalence ratio at lean blowout for the SRL burner was always higher
than that observed for the UDRI burner. The differences were slight, generally on the same order as the
data scatter and measurement uncertainty, but consistently in the same direction on both the SRL and
UDRI test facilities. There was no conclusive explanation for the observed differences.
Demarcation of Lifted Flame Reeion.bv Spontaneous OH uv Emission Measurements.
The following preliminary conclusions have been drawn from the results of the measurements of
spontaneous uv emission at about 308 nm.
81-12
1. The OH uv emission measurements showed a strong change in relative signal intensity as the
fuel equivalence ratio was changed from a very fuel rich case where the flame was clearly attached, to a
very fuel lean case where the flame was very clearly lifted.
2. There existed a marked inflection in the intensity of the OH emission curve that appeared to
separate the regions of the flame where the flame was attached and lifted. This point of demarcation
appeared to mark the location where the flame was fully lifted.
3. The point of demarcation was very insensitive to air flow rate. The value corresponded to a fuel
equivalence ratio of about 1 .06 to 1 .08 over the entire range of air flows tested.
4. The point of demarcation was used to set the fuel equivalence ratios used for other
experiments. Notably, fuel equivalence ratios of about 0.7, 1.06, 1.3, and 1.6 were established to denote
a well lifted flame, the point of demarcation, an attached flame, and a well attached flame respectively.
LDA Velocity Measurements. The following preliminary conclusions have been drawn from the
results of the LDA velocity measurements.
1 . A small recirculation zone exists in the primary fuel tube. This recirculation zone preferentially
attaches to one side of the fuel tube, which cause a regular asymmetry in the velocity profiles in the near
field region. This asymmetry was also seen in cold flow experiments where CO2 was used to simulate the
fuel flowrate. When air or nitrogen was used in the fuel tube, or when the air in the annular air jet was not
flowing, the velocity profiles in the fuel tube were very symmetric.
2. Limited analysis of the velocity in the step region of the combustor has shown the flow to be
very complex. There is a suggestion from the data that there may be multiple vortices in the comer formed
by the side wall and the bluff body.
3. At times, it was difficult to make LDA measurements in the facility because of the unreliability of
the signal processors, and the difficulties in uniformly seeding the flow with the Ai203 seed particles.
4. Velocity measurements at higher air and fuel flow rates, although highly desirable to fully
understand the operation of the burner, were not possible because of the heat load limitations of the
exhaust hood.
Laser Sheet Lighting Imaoino Measurements: The following preliminary conclusions have been
drawn from the results of the laser sheet images taken.
1 The two dimensional images show a complex structure of the shear between the air jet and the
fuel jet and between the air jet and the outer recirculation zone.
81-13
2. The eddy scale in the shear flows appears to be about the same scale as the annular gap
between the fuel tube, and the air tube.
3. The OH images in particular have shown that the flame is rather fragmented. There does not
seem to be a continuous flame sheet. There do appear to be short regions of flame sheet, but in general
they seem detached from other segments of the flame.
V. RECOMMENDATIONS:
Lean Blowout Limit Measurements. The following recommendations might be considered, if
additional, more precise lean blowout data are required.
1 . The flow meters, particularly the fuel flow meter needs to be replaced with a meter of greater
precision. The meter needs to be very carefully calibrated, and should be able to be read to a precision of
0.1 sipm.
2. The fuel flow capacity in the SRL test facility needs to be increased. At present the propane
flow rate is limited to 108 sIpm if a single flow meter/controller is used. If two meters/controllers are used in
parallel, the flow rate can be increased up to about 124 sipm propane.
3. The hood in the SRL test facility needs to have a greater capacity.
LDA Velocity Measurements. The following recommendations might be considered, if additional,
more precise lean blowout data are required.
1 . Replace the exhaust hood with one of higher thermal capacity (A new hood is currently in the
process of being fabricated, and will be installed shortly.)
2. Increase the flow rate potential of the propane fuel system so that measurements can be made
at higher flowrate conditions.
3. Modify the LDA processors to improve their reliability.
4. Incorporate new particle seeders of improved reliability and control.
5. Conduct additional experiments at other fuel equivalence ratios and air flowrates.
Laser Sheet Liohtino Imaoino Measurements: The f.llowing recommendations might be
considered when additional laser sheet lighted images are taken in the burner.
81-14
1 . There is considerable evidence to suggest that the burner flow and flame characteristics have a
considerable three-dimensional character. Consequently, it is suggested that some off axis images be
taken. Illuminating the test section diagonally through the corner windows (i.e. at 45 degrees), with the
CCD camera remaining on the orthogonal coordinate may allow images to be obtained in or at least near to
the comer metal supports.
Suggestions for Follow-on Research. There is considerable work yet needed to fully understand
the operation of this burner. There is a suggestion that the burner has many three-dimensional features
the need to be quantified. Thus, additional off axis two-dimensional imaging measurements, and off axis
LDA velocity measurements are needed. LDA velocity on the diagonal through the corner windows
could also give insight to the flows the exist behind the metal corners. Measurements of gas temperature
using a CARS system are also needed.
The measurements during this period were limited to 1000 sipm air, and about 70 sipm of propane
because of the heat iimitations of the exhaust hood. A hood of larger capacity needs to be installed, so
that higher flowrates can be accommodated. A new exhaust hood has been fabricated, and will be
installed by the end of summer.
Appilcatlon for a Mini Research Initiation Grant. An application for a mini research grant will be
prepared as a result of the summer fellow research program. The proposed project would include the
installation at the Brigham Young University of a burner identical to the one tested in this summer project at
the Aero Propulsion Laboratory, Wright Patterson AFB during this summer fellowship project. The burner
would have been modified to incorporate a swirling burner supplied by Pratt and Whitney Aircraft Co, East
Hartford, Connecticut. This proposed study would allow testing of swirl components in a single near
axisymmetric burner. The Aero Propulsion Laboratory, Wright Patterson AFB has a companion programs,
the Task 100 burner program, and the Task 200 burner program. The Task 100 burner was the burner
investigated during this summer program. The Task 200 burner incorporates four swirling burners in a
linear array, and is designed to burn liquid fuels. The proposed project would use the Task 100
combustor, and a single Task 200 swirling injector. The proposed BYU project would fill a need to
investigate the swirling burner in a single burner configuration. In addition, to basic combustion
measurements, the proposed program at BYU would begin to investigate the formation of NOx pollutants
at the higher fuel equivalence ratios associated with higher performance gas turbines. The proposed mini
grant would involve the installation and checkout of the burner components that would be supplied by
The Aero Propulsion Laboratory, Wright Patterson AFB
REFERENCES
Lewis, B., and G. von Elbe, (1987), Combustion. Flames and Explosions of Gases. 3'''^ Edition,
Academic Press, Inc.
Longwell, J.P., E.E. Frost, and M.A. Weiss, (August 1953), "Flame Stability in Bluff Body
Recirculation Zones". Industrial and Engineering Chemistry. 45(8). PP 1629-1633
Longwell, J.P., and M.A. Weiss, (August 1955), "High Temperature Reaction Rates in Hydrocarbon
Combustion". Industrial and Engineering Chemistry. 47(81. pp 1634-1643
Sturgess, G.J., D.Q. Sloan, A.L. Lesmerises, S.P. Henneghan and D.R. Ballal. (June 1990), "Design
and Development of a Research Combustor for Lean Blowout Studies", 35th International Gas
Turbine and Aeroengine Congress and Exposition. Brussels, Belgium.
81-16
1990 USAF-UES SUMMER FACULTY RESEARCH PROGRAM/
GRADUATE STUDENT RESEARCH PROGRAM
Sponsored by the
AIR FORCE OFFICE OF SCIENTIFIC RESEARCH
Conducted by the Universal Energy Systems, Inc.
FINAL REPORT
AIRCRAFT HVDC POWER SYSTEM - STABILITY ANALYSIS
Prepared by:
K. Sankara Rao, Ph.D. and Allen Olheiser, B. S.
Academic Rank:
Professor
Department and
Electrical and Electronics Engineering Department
University:
North Dakota State University, Fargo, North Dakota
Research Location:
WRDC/POOCl, Wright Patterson Air Force Base
USAF Researcher:
Joe Weimer
Date:
September 7, 1990
Contract Number:
F49620-88-C-0053
AIRCRAFT POWER SYSTEMS - STABILITY
by
K. Sankara Rao
Alan Olheiser
ABSTRACT
Analysis and modeling of aircraft 270V dc electrical power systems are the main topic
of the research project. HVDC at 270 volts has many advantages over the currently
used three phase electrical power systems in an aircraft. There are some problems,
particularly instability in the presence of a constant power load, which are addressed
in this research. Computer models have been developed for the various components
of the HVDC system and constant power load. The analysis using EMTP is included
in this report.
82-2
Acknowledgment
I wish to thank the Air Force Systems Command and the Air Force Office of Sci¬
entific Research for sponsorship of this research. Universal Energy Systems must be
acknowledged for their help and concern in all administrative aspects of this program.
My experience was rewarding because of many influences. Joseph Weimer provided
me with the necessary help, support, encouragement, and an excellent working envi¬
ronment for conducting the research.
82-3
1 TNTRODTTCTTON:
Aircraft Electrical Power Systems, at present use three phase AC systems and research
and development into the possible use of 270V dc as an alternative is underway.
Electrical Laboratory of Aero Power Propulsion Laboratory at Wright Patterson Air
Force Base is very much concerned with computer modeling of HVDC systems. One
of the main problems foreseen is instability when constant power loads are applied to
the HVDC distribution system.
My research interests are in electriccJ power systems and modeling of electrical power
systems. I have used the Electromagnetic Transients Program (EMTP) for modeling
electrical power system components such as, alternators, transmission lines, rectifiers,
inverters etc. and believe that EMTP is an excellent package for analyzing Aircraft
Power Systems. My past experience in modeling three phase 400 Hz distribution
systems in aircraft was a factor in my being assigned for this research.
2 OBJECTIVES OF THE RESEARCH EFFORT:
Electrical load on an aircraft power system usually consists of a combination of light¬
ing and motor loads. In addition there is a considerable amount of dc load in an
aircraft for flight critical computers etc. The main power supply in an aircraft at
present is a three phase supply at 400 Hz. This frequency must remain constant
82-4
in the presence of variable speed turbines which supply mechanical power to the
electrical generators. There are two methods generally used for obtaining constant
frequency
1. Constant Speed Constant Frequency Systems (CSCF): A Constant Speed Drive
(CSD) is used to keep the generator running at a constant speed even while the
input to the constant speed drive varies considerably. The advantage is that
the electrical output from the alternator is at a constant frequency with very
little harmonic content. The main disadvantage of this system is the very high
maintenance required of the CSD.
2. Variable Speed Constant Frequency Systems (VSCF): In this system a variable
speed alternator supplies a variable freque*''/ three phase power. This output
is rectified and inverted to produce constant frequency three phase power. The
main advantage of this system is low maintenance cost. The major disadvantage
is the presence of harmonics in the output due to the inverter operation.
As can be seen above, the three phase ac distribution system has some drawbacks
and research and development is under way for using a 270V dc supply as the main
electrical power supply. In this system, the inverter portion of the VSCF system can
be eliminated and all of the dc loads can be directly applied to it. The low voltage
dc loads at 28V can be supplied by using dc-dc converters. As far as motor loads are
concerned, inverter fed brushless dc motors, inverter fed induction motors or inverter
82-5
fed switch reluctance motors can be used.
My assignment, as a participant in the 1990 Summer Faculty Research Program
(SFRP), was to model a 270V dc power system on a computer and study the sta¬
bility of the system. The Electromagnetic Transients Program (EMTP), which was
developed in the early seventies for Bonneville Power Administration (BPA), is an
excellent tool for analyzing transient behavior of a power system. This package, used
in the present research, is constantly being revised and modified. It has been found
that constant power loads cause instability in HVDC systems and use of a large ca¬
pacitor and harmonic filters eliminate the instability and reduce the ripple current
drawn from the power system.
A more detailed analysis of the HVDC system when motor loads are present should
be undertaken. Since EMTP can be used to model various types of motors, it is
possible to analyze the system under motor loads using EMTP. A proposal for an
AFOSR Mini Grant for continuing this work at North Dakota Slate University is
being submitted.
3 EMTP
The Electromagnetic Transients Program was developed in the early seventies by
Dr. W. Scott Meyer of Bonneville Power Administration. This program is being
constantly updated and has numerous users all over the world. The program is very
82-6
extensive and has all the features necessary for the analysis of Aircraft Power Systems.
Some of the highlights of the program are as follows:
• The various kinds of elements that can be represented are:
1. RLC branches
2. Distributed lines
3. Switches including diodes, thyristors, power transistors and time controlled
switches.
4. Sources
(a) Voltage sources, both ac and dc
(b) Current sources both ac and dc
(c) Voltage and Current sources controlled by any other variable
(d) Alternators with their full representation
(e) DC generators and motors with their full representation
(f) Induction motors with their full representation
(g) Analytical sources
5. Control System Blocks. These blocks can be linear or non linear. They
can be represented by transfer function blocks for linear components. The
inputs and outputs of the control system blocks can be interfaced with the
voltages and currents of the electrical network.
82-7
• The output results are conveniently obtained in grapbiral form on a. CRT or
on any type of plotter or printer. In addition, there is a tabular output which
provides highlights of all the necessary results.
4 CONSTANT POWER LOAD ON AN IDEAL HVDC SYSTEM
As a starting point for this research, an ideal 270V dc source is chosen as the power
system. The series impedance with the ideal source is chosen as a small inductance
in series with a parallel combination of an inductance and a resistance. The series
inductance represents the subtransient reactance of the alternator in the actual HVDC
generating system. The parallel combination of resistance and inductance represent
the time constant and the open loop error. The rectifier filter is represented by a
series RC circuit across the load. The constant power load is represented by a voltage
controlled nonlinear current source such that the product of the current and the
voltage equals the power. The constant power load is a cyclic load occurring 100
times a second with a duty cycle of 0.5. The circuit diagram is shown in Fig. 1.
The simulation results are shown in Fig. 2. Some of the main points of the results
are the following.
1. There is considerable ringing in the voltage across the load.
2. The current drawn from the source has a very high ripple content with a fun-
82-8
damental frequency of 100 Hz.
An analysis of the circuit shows that the system is unstable for constant power loads.
The system can be stabilized by including a large capacitor across the constant power
load. The size of the capacitor depends on the magnitude of the load, the value of the
inductance in series with the source and the resistance in the circuit. The simulation
results, with a 3000 microfarad capacitor in parallel with the load, are given in Fig. 3.
The ringing in the voltage is now absent but the ripple content of the source current
is still high.
The ripple current can be reduced by inserting series LC circuits with resonant fre¬
quencies equal to 100 Hz and its odd harmonics, in parallel with the load. Fig. 4
shows the effect of inserting these harmonic filters.
5 SIMULATION OF A RECTIFIER WITH A CONSTANT POWER LOAD
The second phase of the research consisted of simulating an ideal three phase source
rectified and feeding a constant power load cycling 100 times per second. The circuit
diagram for this system is shown in Fig. 5 and the results are shown in Fig. 6. The
following conclusions can be drawn from the results.
1. The constant power load causes instability and this instability can be eliminated
by inserting a large capacitor in parallel with the load. The ripple content in
82-9
the load voltage is still high. The ctirrent drawn has a high ripple content of
100 Hz and its harmonics.
2. Introduction of harmonic filters as suggested in the earlier section reduces the
ripple content of the voltage and the current drawn from the supply.
In the next phase of the simulation, the three phase voltage source magnitude is
controlled so as to have the voltage across the load to be 270 volts. The overall
control system is of type one.
The results are shown in Fig. 7 and show that the instability can be eliminated by
inserting a large capacitor.
6 BRUSHLESS DC GENERATOR WITH A CONSTANT POWER LOAD
In the next phase of the research, the stability of a brushless DC generator in the
presence of constant power load is studied. As in the earlier cases, the overall system
is unstable under a constant power load. The instability is due to the fact that the
constant power load appears as a negative resistance and the overall impedance as
seen by the generator has a negative real part. Introduction of a large capacitor in
parallel with the load makes the overall impedance to be stable and the overall system
is stable. The value of the capacitor needed to make the system stable depends on
the level of constant power that is applied as the load. The circuit diagram of the
82-10
system is shown in Fig. 11 and the resnlts are shown in Fig. 8 and 9.
7 CONSTANT POWER LOAD
In all phases of the earlier research it was assumed that the load has a characteristic of
demanding constant powt irrespective of the voltage across it. This load is simulated
as a current source whose magnitude is such that the product of the source current
and the voltage across it is the negative of the power demand.
In actual practice, the constant load, as it is assumed is a power conditioner whose
output voltage is maintained constant irrespective of the input voltage. When this
power conditioner is connected to a resistive load, which is constant the output power
is constant. In this phase of the research a buck switching type of regulator is simu¬
lated and used to replace the current sources. The overall system stability is studied.
The circuit diagram of the regulator is shown in Fig. 10 and the results are shown in
Fig. 11.
An examination of the results shows that a, j constant power load, which is derived
from a switching regulator does not pose any problems at all when the input voltage
is derived from ideal sources. The output voltage of the regulator remains constant
while the input voltage is varied within wide limits. The settling time of the transient
response, when the load resistance is changed, is of the order of a tenth of a millisec¬
ond. This makes the circuit fully capable of operating a pulse load at 100 cycles per
82-11
second.
8 RECOMMENDATIONS
The Electromagnetic Transients Program (EMTP) is a very useful tool for analyzing
aircraft power systems and will be used by this researcher in all simulation work.
This simulation will be compared with the simulations using other software where
possible.
The preliminary study shows that a constant power load, when viewed as a resistor
with a hyperbolic v-i characteristic, poses a stability problem for dc systems. However,
when the constant power load is considered as a constant resistive load on a power
conditioner, the dc system seems to behave normally. Further simulation and analysis
should be done to confirm these studies.
A proposal for mini-grant is being submitted to the AFOSR. The work proposed
includes simulation and analysis of an HVDC system with various load configurations
such as motor loads, actuator loads etc. The motor loads considered will be dc motors
and induction motors connected through inverters using MCTs.
82-12
SWITCHING REGULATOR
HflRMONIC riLTERS
»qu»r>ca 0C2 08/30^90
OC StSTCfl
CftTftCnOR ULROSS TIC LOlW
-ICRS
t cmtuLN 1 1 1 i]i« iz . iuoH^ ■
„ 270V oc SiSTCn
CONSinui roHCR tOflO through R POhCR COtOITIOtlCR
1990 USAF-UBS SUMMER FACULTY RESEARCH PROGRAM/
GRADUATE STUDENT RESEARCH PROGRAM
Sponsored by the
AIR FORCE OFFICE OF SCIENTIFIC RESEARCH
Conducted by the
Universal Energy Systems, Inc.
FINAL REPORT
DESIGN OF A DYNAMIC TEMPERATURE MEASUREMENT SYSTEM
FOR REACTING FLOWS
Prepared by:
Academic Rank:
Department and
University:
Research Location:
USAF Researcher:
Larry A. Roe, Ph.D.
Assistant Professor
Mechanical Engineering Department
Virginia Polytechnic Institute and State University
WRDC/POPT
Wright-Patterson AFB, OH 45433
Abdollah S. Nejad
Date;
31 July 1990
Contract No;
F49620-88-C-0053
DESIGN OF A DYNAMIC TEMPERATURE MEASUREMENT SYSTEM
FOR REACTING FLOWS
by
Larry A. Roe
ABSTRACT
A system for the acquisition of spatially and temporally resolved
temperature data in combustion systems was designed. This approach uti¬
lizes a dual-junction thermocouple probe for determination of instanta¬
neous convective characteristics of the junction, necessary for compen¬
sation of the inherently low transient response. A data acquisition
system for coincident recording of velocity with a laser Doppler velo-
cimeter was configured; incorporating the probe, amplifiers, filters,
analog-to-digital converters, and software. An extensive literature
review was conducted, the system designed, requisite hardware specified
and purchasing initiated. Recommendations for an experimental evalua¬
tion of the concept were presented.
83-2
ACKNOWLEDGMENTS
I express appreciation to all those in the Advanced Propulsion Divi¬
sion who contributed to making my stay both personally enjoyable and
technically satisfying, and to the Air Force Systems Command, Air Force
Office of Scientific Research, for providing sponsorship.
83-3
I. INTRODUCTION
The acquisition of spatially and temporally resolved temperature
data in combustion systems is of vital importance for stability charac¬
terization and control, the informed interpretation of velocity data,
and the evaluation of turbulence correlations and compressibility
effects. An application of primary interest is the concurrent acquisi¬
tion of temperature with velocity data acquired via laser Doppler velo-
cimeter (LDV). The LDV provides a sampling of the fluid velocity by
measuring the Doppler shift of laser light scattered from particles in
the flow. Concurrent measurement of temperature at the same spatial
location as the velocity sample is necessary to properly assess the
influence of reaction and compressibility on turbulence, mean veloci¬
ties, and the validity of the LDV measurements themselves.
Data rates must also be sufficient to track fluctuations in the
flowfield, whether they be traditional turbulence, large-scale struc¬
tures, or instabilities. This requires a method capable of providing
thermal data at about 2 kHz for the anticipated application to research
ramjet combustors. Non-intrusive temperature measuring techniques do
not, at present, offer the combination of spatial and temporal resolu¬
tion and high data rates necessary for correlation with LDV data. A
probe measurement is therefore necessary, with thermocouples offering
the best experience base from which to draw. While thermocouples have
been in use for many years, their scientific application to reacting
flows is still in the developmental stage.
Probe techniques pose survivability and system disturbance problems.
Minimal disruption of the mechanical, thermal, and chemical character of
the flow leads to the selection of small devices, which generally have
reduced survivability relative to more robust designs. The required
83-4
compromise leads to the selection of the smallest probe which can sur¬
vive the environment for a practical data-collection time. Even the
smallest physical probe can significantly perturb the flowfield, espe¬
cially in recirculating and swirling flows, so the parameters measured
are not the same as would exist in the absence of the probe. If, how¬
ever, the purpose of the inveatigj,tion is to evaluate compressibility
effects, stress terms, LDV averaging techniques, or general temperature-
velocity correlations, properly configured thermocouples can provide
reliable information if analyzed properly.
Thermocouple voltages give very accurate information about the tem¬
perature of the junction. The relationship between junction temperature
and instantaneous gas temperature requires some evaluation. The primary
considerations are conduction effects between the wire and the probe
support, radiation exchange, catalysis of chemical reactions on the
probe surface, and the transient response of the probe to changing con¬
ditions in the flowfield. Moffat (1962) provides a good introduction to
some of these problems and offers some considerations for dealing with
them.
Briefly, radiation and conduction corrections can often be
neglected, catalysis is minimized by ceramic coatings, and response can
be extended analytically or electronically. Since the physical size and
thermal capacity of even the smallest survivable thermocouple probe pre¬
vent it from responding directly to fluctuations with frequencies above
approximately 200 Hz, response compensation is necessary. This tran¬
sient, real-time, response compensation is absolutely essential for
time-resolved results and requires the most consideration on our part.
For the case of minimal radiation and conduction effects, the response
of the junction will be such that;
83-5
(1)
Tq = Tj + a{dTj/dt)
where T^, is the corrected temperature, Tj is the junction tempera¬
ture, and a is usually referred to as the time constant. The time con¬
stant is the ratio of junction thermal capacity to convective heat flux.
Since this is related to the local instantaneous thermal properties of
the reacting gas mixture, the information required to calculate the time
constant directly is not available. The evaluation of the time con¬
stant, and the implementation of this first-order response compensation,
are topics considered by many researchers.
II. OBJECTIVES OF THE RESEARCH EFFORT
The goal for the summer program was identify a technique for the
acquisition of spatially and temporally resolved temperature data in a
reacting flow, to identify procedures for the application of this tech¬
nique, and to specify hardware isnd software requirements. The ultimate
goal of the research is to integrate these techniques with an existing
laser Doppler velocimeter and to acquire time-resolved, correlated velo¬
city and temperature data in a swirling, reacting flowfield in a dump
combustor. Hardware lead-time restraints were expected to prevent the
implementation of this program during the summer term, so the short-term
emphasis was on system design.
III. LITERATURE REVIEW
An extensive review of thermocouple techniques was conducted during
the term of the Summer Faculty Research Program? this review has been
significantly edited to fit the space constraints of this document. The
primary items of interest are: flowfield type, size of junction or wire,
coating, radiation and conduction corrections, temporal resolution, and
83-6
response compensation. Virtually all the referenced experimental
studies used wires of Pt and/or Pt~Rh.
Fristrom and Westenberg (1965) provide detailed instructions on the
fabrication of small thermocouples and generating silica coatings for
the elimination of catalysis effects. The coating technique involves
passing the junction through a flame containing silicon dioxide gener¬
ated from the combustion of a compound such as dimethyl siloxane. Kent
(1970) reported that silica is not appropriate for temperatures above
about llOOC, as free silicon diffuses into the junction, altering the
properties of the thermocouple. An yttria-beryllia coating was sug¬
gested as an alternate, tested, and found to give good results.
Yule et al. (1978) offer a good discussion of transient compensation
concepts, and provide a correction for non-cylindrical junctions. (Heat
transfer to a cylinder has often been assumed in predicting thermal
response. ) The time constants for the actual junctions were determined
by observing the response to a step-change in temperature by measuring
the decay in measured temperature after a heating current was switched
off. It was stated that this technique will not work unless the thermo¬
couple wires to the junction are much smaller than the probe lead wires.
Wire size and junction size were 25 and 70 microns, respectively; a typ¬
ical time constant was 30 msec. Some reasonable agreement was achieved
between predicted average time constants based on Nusselt number rela¬
tions and measured data. In a propane-air diffusion flame, average time
constants were determined at different locations in the flowfield by
pulsed overheating (again, a response to a known step change in tempera¬
ture) and used to correct the observed temperatures. These thermo¬
couples were apparently not coated. Independent confirmation of the
observed temperatures was not provided. A two-wire technique, with one
83-7
wire constantly heated to provide heat transfer data for response cor¬
rection, was suggested as a future development.
Lockwood and Odidi (1973) measured mean temperatures in a turbulent
diffusion methane flame with a 40 micron thermocouple. No corrections
for radiation or conduction were made and the thermocouple was not
coated. An average time constant for each location in the flame was
determined by an AC excitation method, a typical value was 10 msec.
This time constant was then applied to an on-line compensation circuit
to obtain a corrected voltage. The use of an average, rather than
instantaneous time constant was estimated to generate errors of up to 10
percent in the mean temperature and 20 percent in the instantaneous tem¬
perature.
Lockwood and Moneib (1980) measured the temperature in a nonreact¬
ing, electrically heated free jet with a 12.7 micron thermocouple com¬
pensated for transient response. (They also concluded that the results
of the previous work were not as good as originally claimed.) An on¬
line, digital compensation network was developed; the appropriate ti.me
constant was still an average value determined by a pulsed overheat
method. The time constant was on the order of 15 msec; noise limited
the max frequency response to 5 kHz. Prodigious quantities of results
were plotted, including mean profiles, PDF's, flatness, and skewness,
and were described as "physically realistic."
Heitor et al. (1985) measured temperature simultaneously with LDV
data in premixed, disc-stabilized, natural gas flames. The maximum
error due to radiation was estimated to be lOOK; a specific correction
was not made. The thermocouple output was digitized and stored, with
the compensation being done off-line by determining the gradient dTj/dt
from the stored data. An average time constant was not used, rather, an
83-8
instantaneous time constant was determined based on the measured velo¬
city, physical characteristics of the junction, and a heat transfer
relation for fine wires. This was a good idea, but didn't work very well
in the reacting flows, as the gas properties are not uniquely determined
by the temperature because of large variations in chemistry. Plots of
temperature PDF's near the edge of the flame showed extreme overcorapen-
sation by this method, so the "instantaneous" time constants were all
multiplied by 0.65 (trial-and-error value) to normalize the PDF's to
ambient temperature. The junctions were left uncoated to keep the time
constants as low as possible. Significant build-up of the titanium
dioxide seeding particles on the thermocouple surface caused the
response to degrade rapidly with run time, with LDV data rates of
100/sec being high enough to render the compensation technique essen¬
tially useless. The wire diameters were again 15, 40, or 80 microns.
Yoshida and Tsuji (1978) evaluated the temperature and velocity dis¬
tributions in a premixed propane flame, but not simultaneously. The
thermocouple was 50 micron diameter and uncoated. Transient compensa¬
tion was accomplished in a conventional manner (initially) by determin¬
ing the time constant from an overheat-response method, then incorporat¬
ing this value in a 5 kHz, RC compensation circuit. This gave maximum
instantaneous temperatures higher than adiabatic flame temperature, and
minimum temperatures lower than ambient. To correct this discrepancy, a
series of different time constants was used in the compensation circuit
until a value was found which normalized the output between the ambient
and adiabatic flame temperatures. This value (40 msec) was then used
throughout the flowfield. Some "mismatching" was expected by the
authors from this procedure. Temperature fluctuations of 400C, at
frequencies above 1 kHz, were found.
83-9
Katsuki et al. (1987) developed a linearization technique for time-
response compensation, and evaluated the procedure by rapidly vibrating
the thermocouple junction across the flame front of a laminar diffusion
flame. Both coated and uncoated wires were used. The technique uti¬
lized for compensation essentially determines a film temperature for the
wire and evaluates the instantaneous time constant based on that temper¬
ature. Variations in chemistry are apparently not accounted for. Dif¬
ferences in measured values between coated and uncoated junctions were
ascribed to radiative effects in laminar diffusion flames and catalysis
in turbulent premixed flames. Results from the new technique were com¬
pared to results using a standard, average time constant compensation
technique. Neither procedure was found to give especially impressive
measurements; the authors concluded that simultaneous velocity data were
necessary to properly account for the variations in time constant. It
was also concluded that coating did not adversely affect the determina¬
tion of real-time data; although the time constant was increased due to
the larger diameter, compensation produced the same temperature PDF's as
for the compensated, uncoated, thermocouple.
Lenz and Gunther (1980) used an uncoated, 50 micron, frequency com¬
pensated thermocouple to determine time-resolved temperature in a free-
jet diffusion flame. Compensation was accomplished by determining the
average time constant at each position in the flow by an overheat-
response method, then using the resulting value in an on-line electrical
compensation network. Conduction, radiation, and catalytic effects were
not corrected. Response to 8 kHz was obtained, limited by noise. Sev¬
eral apparent discrepancies in the temperature data were explained by
physical arguments.
83-10
Brum et al. (1983) used a 25 micron, uncoated, compensated thermo¬
couple to acquire temperature data simultaneously with LDV data in a
swirling, reacting flow. The compensation was similar to that of Lock-
wood and Moneib, using an in-situ overheat method to determine an aver¬
age time constant, and on-line electronic compensation of the signal.
The authors indicated that this average time-constant approach was not
optimum, and attributed some apparent discrepancies in the data to the
instrumentation. It was determined that a variation of 10 percent in
the time constant could lead to errors as high as 50 percent in local
heat flux. An alternative technique, utilizing the instantaneous velo¬
city to modify the time constant, was suggested. (This approach was
later used by Heitor et al., with minimal success.) Probe perturbation
effects were also studied, with large variations in velocity discovered
when the probe was inserted, due to both local and large-scale effects.
This is to be expected in elliptic flows.
Farrow et al. (1982) compared CARS data to results obtained from
uncompensated thermocouples in a laminar methane diffusion flame. The
thermocouples were 50 micron diameter and silica coated. Conduction and
radiation losses were accounted for off-line by using the symmetry of
the flame to generate a known environment. The CARS and thermocouple
temperatures were obtained simultaneously, with a "slight" distance
between the measurement regions. Temporally resolved records were not
presented; the two techniques provided good agreement on average temper¬
ature in some regions.
Chandran et al. (1984) measured temperature with a coated, 25 micron
thermocouple, coincident with LDV data, in a premixed turbulent flame.
An average time constant was determined by a cross-spectral analysis
technique, using the response from two closely spaced thermocouples to
83-11
determine the time constants for both. It was claimed that the use of
an average time constant does not introduce significant error in the
temperature results. Probe interference was minimized by adjusting the
probe configuration until the locally velocity closely matched the velo¬
city measured when no probe was in the flow.
In a theoretical investigation, Chandran et al. (1985) used a two¬
time-constant model for a premixed flame with an assumed bimodal temper¬
ature PDF. A linear response was assumed for each of the two parts of
the flow and time constants calculated using assumed flow properties.
For an assumed square wave input, the response of the system was modeled
and the calculated mean temperature was found. Using only a single time
constant generated errors on the order of 10 percent.
Experimental evaluation of these results in a premixed methane-air
flame was conducted with yttria-beryllia coated thermocouples of 25 and
75 micron wire, operated simultaneously. The junction separation was 1
mm. A cross-spectral analysis technique was used to determine the aver¬
age time constant for each junction, assuming the same environment for
both. An RC circuit was tuned to this mean value to accomplish on-line
compensation. Comparison to coincident Rayleigh scattering results
(with about 1 kHz resolution) showed reasonably good agreement for tem¬
porally-resolved temperature, and errors on the order of 20 percent for
the RMS fluctuation and 10 percent for the mean temperature. These
errors were predicted by the response model.
Elmore et al. (1986) conducted a specific program for dynamic tem¬
perature measurement using a dual-junction method with off-line compen¬
sation. The dual junctions were used to determine the response charac¬
teristics of the smaller thermocouple, which was then corrected. Cor-
83-12
rection was not done on a point-by-point basis, but rather in the
frequency domain.
Essentially, the technique determined the convective heat transfer
coefficient as a function of instantaneous frequency, corrected the
measured temperature for convective response in the frequency domain via
extensive use of FFT's, then inverse-transformed (when appropriate) to
get temporal waveforms. Frequency response was limited to about 250 Hz.
The development program was not intended to provide temperatures in
reacting regions, so the assumption of air as the working fluid was jus¬
tified. In this case, the compensation may very well be a function of
frequency only. Changes in local chemistry, and catalytic effects, were
not addressed. For reacting flows, this technique would probably suffer
from the Scune problems observed by Heitor et al. A finite-difference
conduction correction was applied to all data. A Fortran program to
provide for compensation and conduction corrections was developed. The
design and fabrication details for the probe, and initial testing, were
described in an earlier report (Elmore et al., 1983).
Ozem and Gouldin (1989) used an uncoated, 25 micron thermocouple to
determine temperature statistics in a turbulent, premixed, methane-air
flame. Radiation and conduction were neglected, and response compensa¬
tion was conducted off-line. The thermocouple output was amplified,
low-pass filtered at 20 kHz, and digitized at 40kHz for transfer to a
MicroVax II. An average time constant was used, calculated from a Nus-
selt number relation rather than an experimentally determined response.
The bulk of the evidence indicates that time-response compensation
is necessary, at least in turbulent flows, if accurate, temporally
resolved temperature data are to be obtained. The use of an average
time constant does not appear to be sufficient, and using velocity alone
83-13
as an instantaneous "corrector" to the time constant does not provide
much improvement as the response is a function of temperature, velocity,
and chemical composition. A two-wire technique is definitely indicated,
but frequency domain compensation loses the temporal information
required for coincidence with LDV data and still does not account for
chemistry. An instantaneous, real-time correction, based on the differ¬
ential response between two thermocouples of different sizes, is the
most promising approach.
IV. CONCEPT DEVELOPMENT
It will be assumed that conduction along the thermocouple wires is
insignificant relative to convection between the junction and the flow-
field; this has been shown to be valid for wire 1/d ratios greater than
about 50 (Elmore et al., 1983). Transient radiation will also not be
accounted for; the thermocouple response will be;
Tc = Tj + a(dTj/dt). (1)
The time constant is related to the thermal capacity of the junction
and the convective heat flux;
a = ((5jCjdj2)/(4 k Nu), (2)
where Pj is the density of the junction wire, Cj is the specific
heat of the junction wire, dj is the junction diameter, k is the gas
conductivity, and Nu is the Nusselt number. For small wires, the Nus-
selt number will have the form
Nu = CiRe'^-Pr^^, (3)
where Re is the Reynolds number based on wire diameter, and Pr is
the gas Prandtl number. The relations for Nu, Re, and Pr can be substi¬
tuted into the relation for the time constant a to yield
a = ((.jCjdj2-C2)/4) (M^-/{vC^9C^Pr^^Cik)). (4)
83-14
The first grouping of terms is a funct.ion of the junction construction
only, while the second group is related to the instantaneous gas
properties, such than
a = C4G (5)
and
T(, = Tj + C4 G dTj/dt. (6)
For two junctions with differing values of C4 exposed to the same
environment, the instantaneous temperatures Tjj and Tjj^ will be differ¬
ent, out the gas parameter G and actual temperature T^, will be the same.
With known values for the junction properties (?j , Cj , and dj , the values
of C4 are known, and the only remaining unknown is G, which can be
determined;
G = (Tjj - Tjjr)/((C4dTj/dt)j. - (C4dTj/dt)3j) (7)
With the gas parameter known, the time constant for either junction
can then be quantified, allowing a solution for the instantaneous gas
temperature.
It should be noted that this is a dynamic calibration, dependent on
the different responses of the two junctions to the same transient
input. The only required information on the heat transfer characteris¬
tics is the Reynolds number exponent, C2. This value will be about 0.5
(Moffat, 1962), but there is a sensitivity to this assumption. An over¬
heat technique in a nonreacting flow can provide an empirical value for
the exponent.
V. RECOMMENDATIONS
Temperature Measurement System
It is recommended that an experimental verification of this concept
be conducted. An initial program will not require coincident velocity
83-15
measurement, but the data acquisition system should provide for this
capability. The recommended system design point is 1 kHz LDV rate,
3000F gas temperature.
The probe should consist of two junctions, of 25 and 50 micron
Pt/Pt-13%Rh wire (Type R) . The separation of the two junctions should
be approximately 0.5 mm, providing spatial resolution comparable to that
of the LDV system. Lead lengths of 2.5 mm will be sufficient to mini¬
mize conduction effects. The junction wires are to be welded to larger
diameter wires of the seime composition, contained in a protective
sheath. An yttria-beryllia coating should be applied after junction
welding, using the technique described by Kent (1970).
The thermocouple outputs should be amplified and low-pass filtered.
Appropriate amplifier/filter systems are available and quotes have been
obtained. An amplification factor of 100 is needed for 2.0 volts output
at design temperature; the recommended filter cuts are 10 Hz, 500 Hz, 1
kHz, 5 kHz, and 10 kHz.
The amplified output should then be digitized for analysis purposes.
A minimum of three input channels are necessary, two for the amplified
and filtered thermocouple signals, and a third for recording of velocity
from the LDV system. Coordination of the temperature records with data
from the LDV computer is thus possible. A initial data acquisition rate
of 20 kHz is recommended. Approximately 10 to 100 temperatures should be
recorded for each velocity, so that the gradient dE/dt can be accurately
calculated for response compensation.
Compensation can be done off-line, using the known physical charac¬
teristics of the wires and the assumption that the two junctions are
exposed to the same environment. While no papers using such a technique
have been reviewed, the abstract of a Japanese-language paper (Yamazaki
83-16
St al., 1983) appears to describe a similar concept. (A copy of this
paper has been obtained, and translation by USAF agencies has been
requested. ) Since the assumption of identical junction environments is
required, a spectral analysis of the thermocouple voltages, including
coherence and phase information, should be conducted as an indicator of
the maximum reliable frequency response of the system.
Combustors
To properly characterize the capabilities of this novel measurement
technique, tests should be conducted with multiple combustor geometries.
Initial testing should be in a burner providing a laminar premixed flame
and a concentric flow of inert gas. A variable frequency, electro¬
mechanical actuator system may be used to rapidly alternate the probe
between the hot reaction zone and the cold outer flow. This will gener¬
ate a known-frequency transient in the thermocouple output. The chemi¬
cal composition and temperature of the outer flow can be changed to ver¬
ify the convective heat transfer calculations in this region and provide
added confidence in the analysis technique. Traversing frequency can be
varied to identify both the uncompensated and compensated response lim¬
its o? the probe.
Another test sequence should utilize a turbulent premixed flame,
stabilized on a laboratory combustor. The probe would be stationary,
relying on the turbulent fluctuations of the flame front for the tran¬
sient component of the temperature signal. Evaluation of the thermo¬
couple response can be based on observed response from the first test
series and analysis of mean calculated temperatures and temperature dis¬
tribution functions.
83-17
A third test series should then be conducted in a high velocity,
ducted burner. Tests in this combustor will provide information on
probe response and survivability in practical propulsion research sys¬
tems.
After full characterization of this approach to real-time thermal
data acquisition, it is recommended that a direct comparison with avail¬
able non-intrusive diagnostic methods be conducted. For dynamic
measurements, this would require a very clean flame and Rayleigh scat¬
tering apparatus; C2VRS data could be compared at low data rates.
REFERENCES
Brum, R. D., E. T. Seiler, J. C. LaRue, and G. S. Samuelson, "Instanta¬
neous Two-Component Laser Anemometry and Temperature Measurements in a
Complex Flow Model Combustor," AIAA-83-0334, 1983.
Chandran, S. B. S., N. M. Komerath, and W. C. Strahie, "Scalar-Velocity
Correlation in a Turbulent Premixed Flame," 20th Symposium on Combus¬
tion. The Combustion Institute, pp. 429-435, 1984.
Chandran, S. 3. S., N. M. Komerath, W. M. Grissom, J. I. Jagoda, and W.
C. Strahie, "Time Resolved Thermometry by Simultaneous Thermocouple and
Rayleigh Scattering Measurements in a Turbulent Flame," Comb. Sci. and
Tech. , vol. 44, pp. 47-60, 1985.
Elmore, D. L., W. W. Robinson, and W. B. Watkins, "Dynamic Gas Tempera¬
ture Measurement System," NASA CR-168267, 1983.
83-18
Elmore, D. L., W. W. Robinson, and W. B. Watkins, " Further Development
of the Dynamic Gas Temperature Measurement System," NASA CR-179513,
1986.
Farrow, R. L., P. L. Mattern, and L. A. Rahn, "Comparison Between CARS
and Corrected Thermocouple Temperature Measurements in a Diffusion
Flame," AppI. Optics, vol. 21, no. 17, pp. 3119-3125, 1982.
Fristrom, R. M. , and A. A. Westenberg, Flame Structure, pp. 170-174,
McGraw-Hill, 1965.
Heitor, M. V., A. M. K. P. Taylor, and J. H. Whitelaw, "Simultaneous
Velocity and Temperature Measurements in a Premixed Flame," Exp, in
Fluids, vol. 3, pp. 323-339, 1985.
Katsuki, M., Y. Mizutani, and Y. Matsumoto, "An Improved Thermocouple
Technique for Measurement of Fluctuating Temperatures in Flames," Comb,
and Flame, vol. 67, pp, 27-36, 1987.
Kent, J. H., "A Noncatalytic Coating for Platinum-Rhodium Thermo¬
couples," Comb, and Flame, vol. 14, pp, 279-281, 1970.
Lenz, W., and R. Gunther, "Measurements of Fluctuating Temperature in a
Free-Jet Diffusion Flame," Comb, and Flame, vol. 37, pp. 63-70, 1980,
Lockwood, F. C. , and H. A. Moneib, "Fluctuating Temperature Measurements
in a Heated Round Free Jet," Comb. Sci. and Tech., vol. 22, pp. 63-81,
1980.
83-19
Lockwood, F. C. , and A. O. Odidi, "Measurements of Mean, Root-Mean-
Square, and Energy Spectrum of Fluctuating Temperature in a Round, Free,
Turbulent Diffusion Flame", Comb. Inst. European Svmo. . The Combustion
Institute, pp. 507-512, 1973.
Moffat, R. J., Temperature, Its Measurement and Control in Science and
Industry, vol. 3, part 2, pp. 553-571, Reinhold, 1962.
Ozera, H. L., and F. C. Gouldin, "Temperature Time-Series Measurements in
Premixed V-Shaped Flames," Paper 89-36, Eastern Section of The Combus¬
tion Institute, 1989.
Yaroazaki, M. , and M. Ohya, "Measurement of Fluctuatj.ng Temperature in
Turbulent Flame by Dual Thermocouple Method." Nenrvo Kvokai-Shi. vol.
62, no. 673, pp. 318-326, 1983.
Yoshida, A., and H. Tsuji, "Measurements of Fluctuating Temperature and
Velocity in a Turbulent Premixed Flame, " 17th Symposium on Combustion.
The Combustion Institute, pp. 945-956, 1978.
Yule, A. J., D. S. Taylor, and N. A. Chigier, "Thermocouple Signal Pro¬
cessing and On-Line Digital Compensation," J. Energy, vol. 2, no. 4,
pp. 223-231, 1978.
83-20
1986 USAF-UES SUMMER FACULTY RESEARCH PROGRAM/
GRADUATE STUDENT RESEARCH PROGRAM
Sponsored by the
AIR FORCE OFFICE OF SCIENTIFIC RESEARCH
Conducted by the
Universal Energy Systems, Inc.
FINAL REPORT
Hydrogen Permeation in Metals at Low Temperatures
Prepared by;
Academic Rank:
Department and
University:
Research Location:
USAF Researcher:
Date:
Kaveh A. Tagavi
Associate Professor
Mechanical Engineering
University of Kentucky
WRDC/POOS-3
Wright-Patterson Air Force Base
Dayton, OH 45433
Michael J. Morgan/Jerry E. Beam
30 September 90
F49620-88-C-0053
Contract No.:
Hydrogen Permeation in Metals at Low Temperatures
by
Kaveh A. Tagavi
ABSTRACT
Hydrogen permeation through metals at cryogenic to low temperatures is
considered in this report. Unlike permeation in high temperatures, there are very few
data on low temperature permeation. Supercritical hydrogen has been suggested as a
prime candidate for cooling of vacuum tubes aboard spacecraft. The knowledge about
hydrogen permeation at low temperatures, therefore, is essential in establishing
feasibility of this idea. In this project, a comprehensive literature search is conducted in
order to document the state of the art research efforts on hydrogen permeation in metals
at low temperatures. The phenomenon of permeation is investigated and the relevant
parameters affecting it are identified. An apparatus based on vacuum method is
constructed, fabricated, and calibrated. Gathering of actual permeation data are planned
for future activities. As a part of this effort usage of exotic material such as gold,
diamond, or zinc plated metals; graphite copper compounds; and beryllium oxide will be
investigated.
84-2
ACKNOWLEDGEMENTS
I would like to thank Dr. Jerry E. Beam for selecting me for and entrusting me with this
exciting and challenging project. He was always available for consultation and giving me
overall guidance. Michael J. Morgan was there on a daily basis to encourage and support
this project. I always enjoyed and benefited from his positive attitude. My thanks also
goes to Donald Reinmuller for his dedication and constant effort to put life into this
project.
My experience at WPAFB this summer was rewarding. The staff of the Thermal
Technology Group, although diverse in technical expertise, is united on one philosophy
- make the visitor feel like he is at home. They certainly succeeded at this.
I am also thankful to Air Force Office of Scientific Research and the Universal Energy
Systems for sponsoring my appointment.
Finally, my visit was enriched by racquetball games, coffee break discussions, and other
social activities.
I. INTRODUCTION:
Using hydrogen as a coolant in space applications seems both logical and
economical. It’s engineering feasibility, however, is not obvious and has to be
investigated. It is proposed that liquid hydrogen be on board as fuels, which makes it
attractive as a coolant. A hydrogen coolant has two major advantages: first, it does not
add substantially to the overall mass of the spacecraft (since it already exists on board as
fuel) and second, it does not have to be recirculated and cooled since upon absorbing the
thermal load it will be burned as fuel.
The Power Technology Branch of the Aerospace Power Division at Wright
Research and Development Center has been involved in assessing the feasibility of
hydrogen as a coolant for several years. An example of a device that needs to be cooled
is vacuum tubes. However, one of the major concerns regarding cooling of vacuum tubes
by hydrogen is hydrogen mobility through metals. As hydrogen comes into contact with
metal surfaces it starts permeating through them. If enough hydrogen is permeated
across the tube wall and enters into the tube, the ultra high vacuum that exists in the tube
will be compromised. Therefore, it is essential that the magnitude of hydrogen
permeation be estimated. Data on hydrogen permeation are readily available for
common metals at mostly moderate to high temperatures (500-1300 K). But at lower
temperatures of supercritical hydrogen (cryogenic to room temperatures, 35-300 K),
little or no data exist.
My formal education and research interests have been in the general area of heat
and mass transfer. My particular background and interest in space applications and
microgravity-related research contributed to my assignment at the Power Technology
Branch.
II. OBJECTIVES OF THE RESEARCH EFFORT:
Presently, there are no extensive data on hydrogen permeation through metals at
cryogenic to low temperatures." It is the objective of the efforts initiated at the Power
Technology Branch (WRDC/POOS-3) to establish a data-base on hydrogen permeation
at low temperatures and evaluate the feasibility of using supercritical hydrogen as
coolant for removing heat from microwave vacuum tubes. The specific objectives are:
• Perform a literature search and establish the knowledge-base available on the
theories and mechanisms of hydrogen permeation.
• To construct, install, and calibrate a set-up which would measure hydrogen
permeation at cryogenic to low temperatures.
• To establish a base-line permeation data and compare it with the existing data.
84-4
• To examine whether hydrogen permeation levels at cryogenic to low temperatures
are significant enough to interfere with the normal operation of vacuum tubes.
• To investigate whether exotic materials such as gold, zinc, or diamond plated
metals; graphite cooper; and beryllium oxide will inhibit or slow down hydrogen
permeation.
• To recommended directions for future research and development efforts in this
subject.
III. LITERATURE SURVEY:
Hydrogen permeation at high temperatures has been the subject of extensive
research in the past due to its applications in such areas as nuclear reactors (fission and
fusion), metal processing, and electrochemical processes. But only recently research on
low temperature hydrogen permeation has been emphasized due to its applications in
superconductivity and cooling in space. It has been suggested by Chow (1988) that the
already available hydrogen on board as a fuel be used to overcome the thermal load on
spacecrafts. This idea has many potential advantages along with some disadvantages.
The most serious disadvantage is hydrogen permeation.
One of the devices that is proposed to be cooled by supercritical hydrogen is a
microwave tube kept at very high vacuum. While being cooled by hydrogen, some
hydrogen m.olecules may permeate inside the tube and corrupt the vacuum inside. This
will cause malfunctioning of the microwave tube. It is hoped that at low temperatures,
the rate of hydrogen permeation is low enough that a reasonable vacuum can be
maintained during the operation time of such tubes (Pais, 1990).
Hydrogen is very mobile in metals. In fact, its mobility is several orders of
magnitude larger than other interstitials such as oxygen and nitrogen. In addition to
moderate to high temperatures, hydrogen diffusion can also be observed at low
temperatures due to quantum effects at the molecular level (Schaumann et al, 1968;
Wipf and Alefeld, 1974; Rowe et al., 1971).
IV. MECHANISMS OF PERMEATION:
To understand permeation one needs to understand solubility in, diffusion through,
and chemical reaction with the metal in question. Once the metal wall is exposed to
hydrogen, an equilibrium condition between the metal surface and the adjacent
hydrogen is achieved.
Depending on the type of the metal, five types of reactions between hydrogen and
the metal may occur.
84-5
1. Metal groups lA and IIA. Hydrogen forms hydrides. Examples of these metals are:
Na and Ca,
2. Metal groups IVB, VB, and VIB. Hydrogen forms covalent hydrides. Examples of
these metals are: C, Si, S, Se, and As.
3. Hydrogen forms true solutions in some metals. Examples of these metals are: Cu,
Ag, Cr, Mo, W, Fe, Co, Ni, Al, and Ft.
4. Metal groups IIIA, IVA, and VA. Hydrogen forms pseudo hydrides. Examples of
these metals are: Ce, La, Ti, Zr, Th, V, Cb, and Ta.
5. Hydrogen is not absorbed by Au, Zn, Cd, In, or Tl .
The relationship between the adjacent hydrogen partial pressure and the
concentration of hydrogen just inside the metal surface is governed by the well-known
Henry’s law (Johnson, 1988). The higher the hydrogen partial pressure, the higher the
concentration of hydrogen in the solid. Once, a concentration gradient is established,
hydrogen is diffused in the direction of decreasing concentration according to Pick’s law.
Based oi??1^ry’s law, one expects to observe a linear relationship between the
partial pressure of the adjacent hydrogen and the permeation rate Winkelmann was the
first to suspect that the rate of permeation of hydrogen through platinum is not
proportional to pressure (as was originally expected) but rather is proportional to square
root of pressure. His explanation was that the hydrogen molecule first dissociates into
atoms and then diffuses into and through the platinum plate. This theory was later
confirmed by the careful experiments of Richardson et al. (1904) and Richardson (1904).
Besides pressure, hydrogen permeation depends on many other parameters such
as:
• Temperature
• Surface condition
• Stress
• Atomic structure and kinetics
• Chemical reaction between hydrogen and material
In the following sections parameters that affect hydrogen permeation in metals are
described.
V. MATHEMATICAL FORMULATION:
Assume a thin metal membrane is exposed to hydrogen. The temperature and
pressure at each side in general is assumed to be different. As was mentioned before,
the hydrogen molecule enters a chemical reaction (dissociation) in the presence of the
metal and breaks into two hydrogen atoms. The hydrogen atoms subsequently dissolve
84-6
in the metal. According to Henry’s Law the concentration of hydrogen just inside the
metal surface will be
C«Ph (1)
where C is the concentration and Pjj is the pressure of the hydrogen atom gas. But due
to the kinetics of the chemical reaction (dissociation)
Ph.Ph^>/» (2)
and therefore
C(P,T) = S(T)Ph^i/2 (3)
where Pjj^ is the pressure of hydrogen molecule gas and S is the solubility. The above
equation is also known as Sieverts’ law. The parameter S(T) is the constant of
proportionality (indicative of solubility) which is solely a function of temperature and is
independent of pressure and concentration. It is reported by Smithells and Ransley
(1935) that S depends on temperature as
S(T)«Ti/2 exp(-Es/RT) (4)
where T is the temperature, Eg is the energy of solution, and R is the gas constant. Eg is
positive for endothermic reactions such as Hj in copper and negative for exothermic
reactions such as H , in vanadium or niobium. The square term in equation (4) is usually
ignored. It is argued that the temperature dependence of the square term is weaker than
that of the exponential term for moderate and high temperatures. (This might not be as
accurate for low to cryogenic temperatures.) The S expression may now be rewritten as:
S(T) = So exp(-Es/RT) (5)
where is the solubility pre-exponential. Combining these two equations yields:
C(P,T) = So exp(-Es/RT) P^/^ (6)
Hydrogen diffusion (permeation) caused by concentration gradient for a steady state
situation may be written as
J„=DvC (7)
where is the steady state flux and D is the diffusion coefficient. Assuming pressure P
(which results in concentration C) on one side and vacuum (which results in zero
concentration) on the other side
J„ = DC/L (8)
where L is the length along the diffusion path. Substituting for C
J„ = (D/L) So exp(-Es/RT) P^ (9)
It should be mentioned here that the flux is not proportional to the square root of the
pressure difference rather the difference of the square root of pressure. Permeability is
defined as
=4>(T)(Pi/2/L) (10)
84-7
That is
$(7) = DS (11)
where $ is called permeability. The dependence of diffusion coefficient on temperature
is written in terms of the energy of migration
D = D„ exp(-Em/RT) (12)
Accordingly
$(T) = exp(-EjyRT) (13)
where and are diffusion coefficient and permeability pre-exponentials, and Ejn
and Ep are energies of permeation and migration. The energy of permeation Ep may be
expressed as
Ep = Eg + Ejjj (14)
and
*o=DoS„ (15)
It is important to note that $ is only a function of temperature and not pressure.
This is the result of separating the pressure dependence in equation (10). Accordingly
•fo is neither a function of pressure nor temperature and is considered a constant.
VI. FACTORS INFLUENCING HYDROGEN PERMEATION:
There are many parameters that influence permeation. The mechanisms of many
of the dependencies are not well understood. It should be mentioned that most of our
present knowledge is based on experimental observation or on phenomenological
reasoning. This plus the fact that not many experiments have been conducted for low
temperature’ makes the low temperature stipulations risky.
Below, the dependency on some of the more important parameters are discussed.
PRESSURE - Ordinarily it is expected that permeation is proportional to pressure
because of Henry’s law. However, since the hydrogen molecule first dissociates into
hydrogen atoms and then diffuses, permeation becomes proportional to the square root
of pressure. The latter is more or less universally accepted for the case of hydrogen
permeation in metals. One exception is permeation under very low pressures where the
permeation becomes proportional to pressure.
TEMPERATURE -- It is widely believed that tcinperature dependency follows
Arrhenius relationship. However, a single Arrhenius relation might not successfully
represent permeation for a large temperature range of interest -- namely 10 to 1000 K.
Experimental data are available for temperatures down to 0° C (Johnson, 1988). There
are some theoretical work (e.g., Valone et al., 1985), however, that considers permeation
of hydrogen in copper at low temperatures down to 100 K. It was shown that if quantum
84-8
mechanics effects are included, the diffusion coefficient at low temperatures will deviate
appreciably from those calculated based on classical theory. For example, diffusion of
hydrogen in Cu(lOO) at 100 K will be 7200 times higher than that calculated based on
classical theory (Valone et al., 1985).
This deviation at low temperatures has also been observed experimentally for
hydrogen permeation in tantalum (Kokkinidis, 1977) for temperatures less than 200 K
and as low as 100 K and in niobium (Schaumann et al., 1968) for temperatures less than
250 K and as low as 120 K.
One other known effect of temperature in the classical theory (Arrhenius relation)
that has been ignored by most studies is the so called square term shown in equation (4).
Figure 1 shows the ratio of permeability with and without the square term. This figure is
plotted based on preserving the permeability value at the room temperature, hence the
value 1 at 300 K. As seen the effect of this square term in not very large for
temperatures near room temperature. As the temperature decreases, however, this
effect increases. At around 50 K the square term would cut the permeability more than
50%. It should be mentioned that the elimination of this term causes overestimation of
the permeation levels.
The concentration level does not always increase with temperature. If the enthalpy
of solution is negative (due to exothermic solution reaction) then the level of
concentration for a constant pressure decreases as the temperature increases. Similarly,
if enthalpy of permeation is negative, the permeation rate will decrease with
temperature. Figure 2 shows the hydrogen solubility levels for various metals (Fast,
1965). Therefore, in selecting an optimum material, we should avoid materials that react
with hydrogen exothermically, since the cooling with hydrogen will occur at very low
temperatures (as low as cryogenic temperatures).
concentration - According to the simple theory presented here, the diffusion
coefficient does not depend on the level of concentration while the permeation is
proportional to it (see equation (11)). Experimentally, however, some dependence on
concentration has been observed. According to Gissler and Alefeld (1970) values of up
to three times smaller for diffusion coefficient may be observed for higher
concentrations of hydrogen permeating in niobium.
At high concentrations of hydrogen there is an overshoot effect (Beck et al., 1966).
For example for a hydrogen-iron system, the permeation rate was observed to reach a
maximum value and then decay to a steady state value. The overshoot was as much as
100% of the steady state value and lasted for several minutes.
SURFACE CONDITION - Hydrogen permeation depends strongly on the surface
84-9
condition. Both the macroscopic and microscopic structure of the surface as well as the
presence of oxides affect hydrogen permeation. The most severe problem with gas
phase charging is the presence of oxides at the surface which are presumed to hinder
hydrogen entry or exit at temperatures about 300 K (Kumnick and Johnson, 1975) . This
explains the fact that most charging experiments are performed for temperatures well
over 300 K. It has been shown (Kumnick and Johnson, 1975) that this difficulty can be
overcome by deposition of a thin layer of palladium on each surface of the permeation
membrane. The palladium layer acts as an excellent catalyst for hydrogen dissociation.
With this technique gas phase permeation methods mav be carried out at temperatures
well below 300 K
STRESS " Stress has been shown not to affect the diffusion coefficient, but it increases
the solubility of hydrogen in the lattice (Beck et al., 1966). As a result, permeability,
which is equal to the product of diffusion coefficient and solubility, will be increased
under stress. The mechanical behavior of metals in the presence of hydrogen is further
investigated by Toriano (1960).
VII. EXPERIMENTAL METHOD:
There are several methods for measuring diffusion (permeation) of hydrogen in
solids. Volkl and Aiefeld (1979), for example, have given details of all these techniques.
One direct way of measuring permeation is a technique whereby a thin membrane of the
metal of interest is exposed to pure hydrogen on one side and vacuum on the other. As
hydrogen permeates through the membrane, the pressure on the vacuum side increases.
By measuring the pressure variation with time, information about the diffusion
coefficient, solubility, and permeability is obtained.
Permeation measurements may be performed in two different ways: stationary’ and
time-lag methods (Volkl and Aiefeld, 1979). The stationary method reveals the product
of the diffusion coefficient and the solubility. The time-lag method reveals both
diffusion coefficient and solubility. The diffusion coefficient can be deduced from the
relaxation time while the solubility data can be determined from the change in
permeation rate.
At steady state, permeation rate is proportional to the product of concentration and
diffusion coefficient (Barrer, 1951). While transient data reveals trapping parameters in
addition to concentration and diffusion coefficient (Johnson, 1988). It is reported by
Johnson (1988) that the time lag decreases with temperature and pressure. For hydrogen
permeation in iron, time lags between 1000 to 10,000 seconds were observed for a
sample of about 0.8 mm at standard temperature and pressure.
84-10
The present set-up utilizes the vacuum technique. Figure 3 shows the schematic of
the set-up. It essentially consists of two chambers. One contains hydrogen at slightly
above atmospheric pressure while the other is maintained at ultra high vacuum (between
1 X 10*® and 1 x 10*® torr). The two chambers are separated by a copper disk, 0.080 inch
(about 2.03 mm) thick, which also serves as the test specimen. All joints are sealed by
rotatable flanges that utilize knife-edge technology. This will provide seals that are
appropriate for maintaining ultra high vacuums. The desirable vacuum is reached by two
pumping units - a molecular pump and an ion pump. The molecular pump evacuates
the chamber at pressures near the atmospheric levels down to 1 x 10*® torr levels. At
this point, the pumping action is switched to the ion pump which is capable of creating
vacuums of up to 1 x 10"^ ° torr range. The pressure in the vacuum side is measured with
a Convectron gauge (for low vacuum) and a Nude gauge (for ultra high vacuum).
The hydrogen and vacuum chambers are wrapped with heating tapes and cooling
coils. Two options may be used as a liquid coolant inside the coils -- liquid nitrogen or
coolant derived from an ultra low temperature (down to -90* C ) circulating bath. The
combination of the heating tapes and the cooling coils allows us to maintain any desired
specimen temperature in the range -150* C to 250* C. The whole assembly including the
pump is wrapped with thick insulation to facilitate temperature control.
VIII. EXPERIMENTAL PROCEDURE:
Before a permeation test can be performed, there are a couple preparatory tests
that must be conducted: 1) leak detection, 2) system bake-out.
First, the vacuum chamber must be tested against leaks. One way this test is
conducted is by turning off the vacuum pump after reaching an ultra high vacuum and
observing the system pressure. A typical pressure-time history is shown in Figure 4.
After the pump is turned off, the pressure starts to rise for two distinct reasons -- leaks
and outgassing. (Outgassing is the diffusion of dissolved gasses out of the chamber inner
surfaces and into the vacuumed space. The driving force for outgassing is the difference
between the chemical potentials of the dissolved gas and of the same gas in the chamber.
That is, the lower chamber pressure the higher the rate of outgassing.) If there is a leak,
the pressure ultimately equals that of the surrounding atmosphere. In the absence of any
significant leak, the pressure in the vacuum chamber rises to a unique pressure
(equilibrium pressure) corresponding to the amount of gases dissolved at the inner
surfaces of the vacuum chamber. After this pressure is reached, the rate of outgassing
quickly drops to zero and no further significant increase in the chamber pressure is
observed. In order lo lower the equilibrium pressure, a procedure called baking-out is
84-11
employed.
The chemical potential of the dissolved gases increases drastically with
temperature. Therefore if the chamber is heated and vacuumed simultaneously,
considerable amounts of dissolved gases may diffuse out and be taken out through the
vacuum system. After the system is cooled, a lower level of equilibrium pressure may be
attained due to the decreased level of dissolved gases. Baking periods of up to several
days at elevated temperatures of 150-250“ C are recommended in order to obtain ultra
high vj cuums.
After the leak detection and baking out the set-up is ready for permeation tests.
The principle tenet of the experiment is to measure the time (At) it takes for the
chamber pressure to rise from the initial pressure of Pj (it is best that this pressure be
greater or equal to the equilibrium pressure) to the final pressure of Pf. Assuming that
the pressure of the hydrogen chamber (Ph^) is much greater than Pf and that it is held
constant during the experiment, and using equation (10), the following relationship
between the diffusion rate and the pressure buildup inside the vacuum chamber may be
written;
t = lV(Pf.Pi)LWRT Pu, Pstpl (16)
where V is the volume of the vacuum chamber, A is the area of the test specimen
available for permeation and Pstp is the density of hydrogen at standard temperature and
pressure. Figure 5 shows the amount of time necessary for the pressure rise from 1 x
10“® to 1 X 10"* torr for our present set-up with copper as the specimen. As seen, at
room temperature, it takes about 2.15 hours for the pressure buildup. However, for
lower temperatures, the time required increases rapidly. For example, at 200 K, it takes
about 100 million hours for the same pressure buildup. It should be mentioned that this
figure is based on Arrhenius relation and that this relationship might not hold for lower
temperatures.
IX. MATERIALS OF INTEREST:
A few materials have been selected for testing in this work each for a specific
reason.
• Copper is selected for the baseline material. The set-up will be checked and
calibrated against data available in the literature.
• Diamond is selected with hopes that its compact and uniform molecular structure
would inhibit hydrogen permeation,
• BeO is selected because of its excellent thermal, structural, and electrical
properties.
84-12
• Graphite compounds such as GrCu is selected for their excellent structural
properties.
In the following sections the available data for each of these materials are listed.
COPPER — There are several references dealing with hydrogen permeation through
copper. One of the earliest, yet widely referred, is by Smithells and Ransley (1935).
Table 1 lists the diffusion data for hydrogen-copper system as well as other permeation
data. More results are shown in Table 2 taken from Dushman (1962). This table lists
the hydrogen solubility data for various temperatures. Alloys of copper have received
some attention too. It has been shown that the hydrogen solubility varies as a function of
the amount of alloying metal. It is seen that the effect of the alloying covers a wide range
of solubility enhancement for Cu-Ni to decrease in solubility for Cu-Al. Figure 2 shows
the solubility data vs. temperature for Cu as well as other metals.
The diffusivity of hydrogen in copper at low temperatures was studied by Ishikawa
and McLellan (1985). The data presented are for temperatures as low as 300 K. Figure
6 show these results.
DIAMOND - No data was found for solubility, diffusion, or permeation of hydrogen
through diamond. However, a point of concern should be any possible reaction of
hydrogen and diamond (carbon) which may result in methane (CH^ ). The intensity and
severeness of such reaction (if any) should increase with temperature. This in turn
increases the importance of predicting hot spots.
BeO - No data was found for solubility, diffiision, or permeation of hydrogen in BeO.
GRAPHITE COMPOUNDS - No data was found for solubility, diffusion, or
permeation of hydrogen through graphite compounds such as Gr-Cu. However, a few
reports are available for graphite (e.g. Kiyoshi et al., 1988; Karimi and Vidali, 1989).
X. CONCLUDING REMARKS/FUTURE DIRECTIONS:
Creating, maintaining, and measuring ultra high vacuums (e.g., 1 x 10"® torr) are by
no account a trivial matter. The ten week period fell just short of obtaining accurate and
repeatable hydrogen permeation data. As a result, this report is focusing on theoretical
background, explaining the physical phenomena, and exploring future directions. Plans
are undergoing for obtaining low temperature permeation data in the future. Below,
some specific remarks and suggestions are listed.
• Since we are dealing with very large time intervals (days or even weeks), the design
should be modified so that larger hydrogen pressures and thinner test specimen
could be used.
• The experimental set-up and procedure should be tested against known data.
84-13
Choices of pure copper or palladium (a precious metal) at or above room
temperature seem appropriate.
• In choosing the material, we should exercise caution in selecting those which react
with (or dissolve) hydrogen exothermically. The solubility of these materials
increase with deo’easing temperatures.
• It is not clear whether BeO reacts with hydrogen endothermically or
exothermically.
• It is reported that certain metals (i.e., Au, Zn, Cd, In, and Tl) do not absorb
hydrogen. Therefore, these metals might be used as barriers to hydrogen
permeation. For example, gold plating a sample might inhibit hydrogen
permeation to a considerable degree.
• Use of diamond coating has been suggested as a hydrogen permeation barrier. For
testing purposes, a metal with a large hydrogen absorption should be coated with
diamond. Since the metal virtually presents no resistance to hydrogen permeation,
the experimental data would reveal permeability of the diamond coatings.
• The literature is scarce when it comes to permeation data for temperatures lower
that 300 K. This plus the fact that the mechanism of hydrogen absorption and
permeation is not fully understood (even for higher temperatures) suggest that we
need data for the actual temperature of operation. In short, data for the range of
room temperature down to 250 K might not necessarily predict data for 200 K or
100 K, One purely theoretical paper suggests that the diffusion coefficient of
hydrogen in copper when including the quantum mechanics effects will be 7200
times higher than that calculated according to the classical (Arrhenius)
relationships.
« Even if experimental data show that hydrogen permeation levels are not significant
at supercritical hydrogen temperatures (about 40 K), we should try to design the
system for higher temperatures. The supercritical hydrogen is cooling a surface
that might be very hot (several hundred Kelvin). Any disruption or instability in
the cooling mechanism might create temporary hot spots in the vacuum tube wall.
These hot spots will act as open windows for hydrogen permeation. This scenario
should encourage the search for material which are more impermeable to hydrogen
(such as gold or diamond plated metals?).
84-14
REFERENCES
Barrer, R. M., Diffusion In and Through Solids, Cambridge University Press,
Oxford University, England, 1951.
Beck, W., Bockris, J. O’M., McBreen, J., and Nanis, L., "Hydrogen Permeation in
Metals as a Function of Stress, Temperature and DissoNed Hydrogen Concentration,"
Proceedings, Royal Society, London, Series A, Volume 290, 1966, pp. 220-235.
Chow, L. C, "Forced Convection Cooling of Microwave Tube Collector with
Supercritical Hydrogen," Final Report SCEEE Contract F30602-81-C-0193, Task 0091,
1988.
Dushman, S., Scientific Foundations of Vacuum Technique, Second Edition,
Edited by J. M. Lafferty, John Wiley, New York, 1962.
Fast, J. D., Interaction of Metals and Gases, Volume 1, Thermodynamic and Phase
Relations, 1965.
Gissler, W., Alefeld, G., Springer, T., J. Phys. Chem. Sol., Volume 31, 1970, p.
2361.
Ishikawa, T. and McLellan, R. B., "The Diffusivity of Hydrogen in Copper at Low
Temperatures," J. Phys. Chem. Solids, Volume 46, Number 4, 1985, pp. 445-447.
Johnson, H. H., "Hydrogen in Iron," Metallurgical Transactions B, Volume 19B,
October 1988, pp. 691-707.
Karimi, M. and Vidali, G., "The Adsorption of Hj, Dj and Ar on Graphite: New
Theoretical Results," Surface Science, Volume 208, 1989, pp. L73-L79.
Kiyoshi, T., Namba, T., and Yamawaki, M., "Hydrogen Permeation through
Graphite," Journal of Nuclear Materials, Volumes 155-157, 1988, pp. 230-233.
Kokkinidis, M., Diploma Thesis, Technical University Munchen, Germany, 1977.
Kumnick, A. J. and Johnson, H. H., "Steady State Hydrogen Transport through
Zone Refined Irons," Metallurgical Transactions A, Volume 6A, 1975, pp. 1087-1091.
Pais, M. R., "Permeability of Hydrogen in Silicon," SBIR Proposal submitted to and
funded by U.S. Department of Defense, 1990.
Richardson, O. W., "The Solubility and Diffusion in Solution of Dissociated
Gases," Philosophical Magazine, Sixth Series, Volume 7, 1904, pp. 266-274.
Richardson, O. W., Nicol, J., and Parnell, T., "The Diffusion of Hydrogen through
Hot Platinum," Philosophical Magazine, Sixth Series, Volume 8, Number 43, July 1904,
pp. 1-29.
Rowe, J. M., Skold, K., Flotow, H. E., and Rush, J. J., J. Phys. Chem. Sol., Volume
32, 1971, pp. 41-54.
84-15
Schaumann, G., Volkl, J., and Alefeld, G., "Relaxation Process due to Long-Range
Diffusion of Hydrogen and Deuterium in Niobium," Physical Review Letters, Volume
21, 1968, pp. 891-893.
Smithells, C. J. and Ransley, C. E., "The Diffusion of Gases through Metals,"
Proceedings, Royal Society, Series A, Volume 150, 1935, pp. 172-197.
Troiano, A. R., "The Role of Hydrogen and other Interstitials in the Mechanical
Behavior of Metals," Transactions ASM, Volume 52, 1960, pp. 54-80.
Valone, S. M., Voter, A. F., and Doll, J. D., "The Isotope and Temperature
Dependence for Self-Diffusion for Hydrogen, Deuterium, and Tritium on Cu(lOO) in the
100-1000 K Range," Surface Science, Volume 155, 1985, pp. 687-699.
Volkl, J. and Alefeld, G., "Hydrogen Diffusion in Metals," in Diffusion in Solids,
edited by Nowick, A. S. and Burton, J. J., Academic Press, New York, 1979, pp. 231-302.
Winkelmann, A., Drude’s Ann. Volume Vtll, p. 388.
Wipf, H., and Alefeld, G., Phys. Stat. Sol. (a), Volume 23, 1974, pp. 175-186,
84-16
Temperacure, K
Figure 1. Effect of term on permeability
Figure 1. Solubility data for various
Table 1. Diffusion data for hydrogen and
various metals from Smithells
and Ransley (1935). D is in
STP cm^ per second through 1 cm^
of surface and across 1 mm thickness,
p is in mmHg.
(1)
(2)
(3)
(4)
System
b
k
Authors
H,
Ni
7,710
2-3 X 10-*
Lombard.
6.930
0-85
Deroing and Hendricks.
6,900
1-4
Borelius and Lindblom.
6,700
I'OS
Ham.
H,
Pt
9,800
1 -41 ;< IO-‘
Richardson.
9,000
1-18
Ham.
H,
.Mo
10,100
0-93 :< 10-*
Smithells and Ransley.
H,
Pd
2,100
41 X 10-*
Lombard and Eichner.
H,
Cu
8.300
0-91 < 10-*
Smithells and Ransley.
H,
Fe
4.800
1-63 -< 10-»
Smithells and Ransley.
4,700
1-60
Borelius and Linablom.
0,
Ag
11.300
3-72 X 10-‘
Spencer.
11.300
2'06
Johnson and Larose.
N,
Mo
21200
8-3 X 10-*
Smithells and Ransley.
N,
Fe
11.900
1-5 X 10-*
Ryder.
CO
Fe
9.300
1-3 X 10-*
Ryder.
Table 2. Hydrogen solubility in copper from
Dushman (1962). Vq = cm^(STP) per I g
metal, s = cm^(STP) per 100 g metal,
Vg = volume of gas (STP) per 1 volume
or metal, v = number of atoms per
atom of metal.
r C:
400
300
600
700
800
900
1000
S'.
0 06
0.16
0.30
0.49
0.72
1.08
1 38
0 0054
0.0143
0.0268
0.0439
0.0644
0.0967
0 141
!0*/-
O.0S3
0.0908
0.170
0.278
0.409
0.613
0.896
CC.
1083 (mpj
1100
1200
1300
1400
1300
S‘.
2.10 (i)
6.00(0
6.3
8.1
10.0
11.8
13.6
0.188
0 337
0.364
0.725
0.893
1.05
1.22
m-
1.19
3.40
3.38
4.60
3.67
6.70
7 71
• Values of V, ana r calcuiaied by Oushman.
84-19
/t»cc
2
3
lOOO/T, K-1
Figure 5
Estimaced time it takes for the present
set-up to go from 10“^ torr to 10”5 corr
— T ‘K
000 ?00 5C30 400 300
Figure o. Diffusivity of copper
at low temperatures
from Ishikausa and
''.cLeilan ■ 1985 > .
84-20
1990 USAF - UES SUMMER FACULTY RESEARCH PROGRAM
GRADUATE STUDENT RESEARCH PROGRAM
Sponsored by the
AIR FORCE OFFICE OF SCIENTIFIC RESEARCH
Conducted by the
Universal Energy Systems, Inc.
FINAL REPORT
MEASUREMENTS OF DROPLET VELOCITY AND SIZE
DISTRIBUTIONS FOR A PRESSURE/AIR BLAST ATOMIZER
Prepared by:
Academic Rank:
Department and
University:
Research Location:
USAF Researcher:
Date:
Contract No:
Richard S. Tankin
Professor
Mechanical Engineering Department
Northwestern University
WRDC/POSF
Wright Patterson Air Force Base
Dayton, Ohio 454336563
Thomas Jackson
21 August 1990
F49620-88-C-0053
MEASUREMENTS OF DROPLET VELOCITY AND SIZE
DISTRIBUTIONS FOR A PRESSURE/AIR BLAST ATOMIZER
by
Richard S. Tankin
Abstract
A phase doppler instrument was used to measure droplet sizes and
velocities in a water spray. This nozzle consisted of a hollow cone water spray and
two swirling, concentric air channels. Three different water flow rates were
examined; and three different air flows. Horizontal traverses were made across
the spray near the sheet break-up region. More than 200,000 samples were taken
in each traverse. The results show that the spray is axially symmetric which is
important for the planned theoretical analysis. The analysis of the data will follow
the same procedure that was developed to analyze the data that was collected last
summer.
85-2
Acknowledgements
1 wish to thank WRDC/POSF at Wright Patterson Air Force Base and the Air
Force Office of Scientific Research for the sponsorship of this research. I also
want to thank Universal Energy Systems for their efficient handling of all
administrative aspects of the program.
This summer's research was beneficial to me and hopefully will lead to joint
research activities with Wright Patterson. Having worked before with Dr. W. M.
Roquemore and Dr. T. Jackson, it was no surprise to find them very cooperative
and encouraging. The experimental work could not have been accomplished
without the expertise and experience of Dr. G. Switzer. The ease and readiness
with which Jeff Stutrud handled complex computer problems was invaluable. 1
found the staff at Wright Patterson to be friendly, stimulating, and helpful.
85-3
L INTRODUCTION
Gas turbines in nearly all cases utilize liquid fiiels. These fuels must be delivered as
small droplets to the combustion zone if efficient combustion is to occur. The
process by which these small drops are formed is called atomization. The method
of achieving atomization is varied and not a topic for consideration in this study. In
this study the droplet distributions from a pressure/air blast atomizer are
examined. The liquid spray discharges from the nozzle as a liquid sheet which
breaks up downstream to form droplets.Surrounding this liquid sheet are two
concentric, swirling air flows. Pictures of the spray from this nozzle were taken
and a typical photograph is shown in Figure 1. The light source for this photograph
was a strobe (pulse duration is about 100 psec).
It is important to determine the spray characteristic if one hopes to correlate
different sets of experimental data, compute numerical simulations, determine
droplet trajectories, etc. In the past, these characteristics were limited to drop size
distribution, pattemation, cone angle, dispersion, and penetration. For example
various techniques have been used to determine the drop size distribution -
photographic, optical, collection devices, etc. Each of these techniques have very
limited capabilities. The Malvern technique obtains a size distribution that is
integrated over the optical path length. To determine radial distribution requires
the use of Abel inversion. Dense sprays, or asymmetry can complicate this
technique. Two situations have recently arisen - a vast improvement in
instrumentation and a new approach for predicting droplet distributions.
la. INSTRUMENTATION
Recently a highly sophisticated instrument has been developed by Aerometrics
which is capable of optically measuring the sizes and velocity of droplets. This
method utilizes light scattering techniques, and consists of a transmitter, receiver,
signal processor, computer ah laser light source. The transmitter has a beam
expander which reduces the size of the measuring point - which in our case is
85-4
about 1 X 10‘4cm2, Thus excellent spatial resolution is achieved. Since the
detectors in the receiver unit are photomultipliers (three), the response time of this
instrument is very short. Signals from individual drops can be processed and the
data transferred to computer memory in 20 psec. For the operator, an important
aspect of this system is the software program associated with the signal
processor. As data are being collected and stored by the computer, they are
presented in bar graph form for the operator to observe. After a selected number
of droplets are collected, a listing of pertinent data - such as attempts, validations,
corrected count, probe area, etc. - are displayed. The bar graphs consist of
particle size counts, and velocity distributions. A typical example of such a display
is shown in Figure 2.
Last summer at Wright Patterson Air Force Base , I collected similar data from a
pressure atomizer. These results were presented at the 1990 Institute for Liquid
Atomizarion and Spray Systems in Hartford, Conneticut (Li, Chin, Tankin,
Jackson, Stutrud, and Switzer). This summer we extended the work to include the
effects of high velocity, concentric, swirling air. To achieve our goal, nine different
conditions were examined - three different water flow rates and three different air
flow rates. I might add that this is a very complicated flow and if we can
successfully predict the size and velocity distributions for this nozzle, I think we
can analyze any nozzle.
Ib. THEORY
The concept of information entropy was developed by Claude Shannon (1948),
and Jaynes (1957) later extended this concept into the well-known method of
maximum entropy formalism. This formalism can be applied to problems involving
probability, i.e., where insufficient information is available to obtain exact solutions.
We have applied the maximum entropy formalism to liquid sprays in which we
predicted the droplet size and velocity distributions. Since the application to this
problem has been adequately discussed by several researchers - Kelly (1976),
Sellens and Brzustowski (1985, 1986), Sellens (1989), and Li and Tankin (1987,
1988, 1989), it will not be necessary to develop the background material once
again ( for details, see Li 1989). The data collected in this study will be examined
using the maximum entropy principle. There has only been two papers in which
comparisons between theory and measurements exist. In one, Li and Tankin
85-5
(1988) used data that were obtained from holographic and photographic methods.
Thus these results were limited to size distributions. The other paper is by Sellens
(1989) in which he use a phase doppler instrument to obtain the data. However in
this study there are some inconsistencies and the experiments are questionable
(see Li and Tankin, 1989). Thus it was necessary to carefully obtain correct data
with which to compare the theory. That is the purpose of this study.
II. OBJECTIVES OF THE RESEARCH EFFORT:
Currently, there are very few precise measurements of radially integrated droplet
size and droplet velocity distributions for sprays. Our goal is to obtain such
measurements under these various flow conditions and to present them in a two-
dimensional matrix form - one axis being the droplet diameter, and the other being
the droplet velocity. From such a two-dimensional matrix of the data, one can then
obtain a variety of other representations - droplet size distribution; velocity
distributions for different diameter particles; total number of droplets; flow rates;
average velocity as a function of drop size, etc. These measurements should be
made as close to the sheet break-up region as possible.
My task this summer was to work with the staff at Wright Patterson Air Force
Base making radially-integrated spray measurements using a phase doppler
particle analyzer and to obtain the data in a two-dimensional matrix form. Then to
extract from this matrix the various probability distribution functions, average
droplet velocity as a function of droplet diameter, flow rate, etc.
These experimental data are important because they can be used to extend the
applicability of recently developed theoretical predictions. One problem in
collecting such data is the requirement that samples be sufficiently large for the
statistical quantities to be significant. In our case we collected more than 200,000
samples per run. Nine such runs were completed.
III. RESULTS
We measured the radial droplet population for three different water flow rates -
4.2 ml/sec, 7.6ml/sec and 8.33ml/sec. The three air flow rates, for each water flow
rate, were zero, l,833ml/sec, and 3,000 ml/sec. When the water flow rate is
85-6
4.2ml/sec, measurements were made in a horizontal plane located 10 mm from
the nozzle exit; when the water flow rate is 7.6ml/sec, measurements were made
7mm from the nozzle; and when the water flow rate is 8.3 ml/sec, measurements
were made 5mm from the nozzle.The radial population was obtained from
measurements at 0.5mm increments over the radius of the spray. The velocity
measurements and Sauter mean diameter measurements indicate the spray is
axially symmetric (see Figure 3). It should be added that during each of these
traverses it was necessary to shut down the spray at least once, refill the
reservoir, and reset the flow rate. Although there may be some variations in
reestablishing the flow, we did not detect it once the flow was restablished. Since
there are two air passages, we needed to know the ratio of the air flow through
each passage. To accomplish this, we blocked each air passage separately and
measured the pressure drop for various air flow rates. The results are seen in
Figure 4 - which indicate the air flow is equally divided between the two passages.
»
A joint size-axial velocity distribution function will constmcted from the individual
point measurements, weighing each measurement by their time of collection and
the ratio of their optical probe area to the ring area represented at that location.
This experimentally determined joint distribution function will then be compared to
that predicted by the maximum entropy analysis.
The experimental data will be similarly handled as the data obtained last summer.
It will be necessary to make some assumptions regarding the source terms in the
theory. The mass source term - Sm - will be assumed to be zero. The other two
source terms - momentum (Smv) ttnd energy (Se) - will require estimates after
examining the experimental data that were collected.
IV RECOMMENDATIONS
The necessary experimental data have been collected and stored on floppy disks.
It may be necessary to obtain from Aerometrics additional information that will
allow us to read these data and calculate the probe area for those runs in which
this information is lacking. Then the integrated two-dimensional matrix will be
formed. Once we have the experimental data in this fonn, the various distributions
and mean values can be determined. An additional quantity that needs to be
85-7
compared is the flow rate obtained from the droplet measurements with the flow
collected in a graduate over a period of time.
Since this nozzle was supplied to us by Allison Gas Turbine Division; we will
report to them, through Wright Patterson, our results. Hopefully similar
experiments will be conducted on a combusting flow using this nozzle. Once these
data are collected, analysis can begin on the source terms for such a flow. In this
flow there will be heat, mass, and momentum transfer - much different from the
cold flow experiments conducted this summer.
VI. REFERENCES:
1. Shannon, C. E., "A Mathematical Theory of Communication ", The Bell System
Technical Journal. 27. 1948, p.379.
2. Jaynes, E.T., "Information Theory and Statistical Mechanics", Physical Review.
106, 1957, p. 620.
3. Kelly, A. J., "Electrostatic Metallic Spray Theory", J. of Applied Physics. 47,
1976, p, 5264.
4. Sellens, R. W. and Brzustowski, T. A., "A Prediction of Drop Size Distribution in
a Spray from First Principle", Atomization and Sprav. 1, 1985, p. 89.
5. Sellens, R. W. and Brzustowski, T. A., "A Simplified Prediction of Droplet
Velocity Distributions in a Spray", Combustion and Flame. 65, 1986, p. 273.
6. Li, X. and Tankin, R.S., "Droplet Size Distribution:A Derivation of a Nukiyama-
Tanasawa Type Distribution Function", Combustion Science and Flame. 56, 1987,
p. 65.
7. Li, X. and Tankin, R.S., "Derivation of Droplet Size Distribution in Sprays Using
Information Tlieory", Combustion Science and Flame. 60, 1988, p. 345.
85-8
8. Li, X. and Tankin, R.S., "On Prediction of Droplet Size and Velocity Distributions
in Sprays Through Maximum Entropy Formalism", Combustion Science and
Technology^ 68 , 1989, p. 147.
10 Li, X., PhD Thesis, Department of Mechanical Engineering, Northwestern
University, June, 1989.
11. Li, X., L.P. Chin, R.S. Tankin, T. Jackson, J. Stutrud, and G. Switzer,
"Comparison Betwwen Theory and Experiments for Sprays Fiom a Pressure
Atomizer", 1990 ILASS Meeting in Hartford, Conneticut.
b
Figure la. This is a photograph of target, where each small division
is 1mm. Figure lb is a photograph of the spray when the water
flow rate is 8.3tnl/sec.
85-10
Ile*n 3,2 itH
2**" <028>= 4.0 u«
S**” <D30)= 3,3 uM
S«ut*r H*«n <D32>= 9.4 uM
Prob« flgj* = 3.33E-003 cm2
= 4.19E*003 /’cc
u2!»«*f?u2 *• = i«g4E-007 oo/s -
VolUM# r*«X = S.S3E-O0S oo/*/om2
17.4 33.8
Di«M«t*x« uM
Attefipts
Ual i<i«tlqng
CoprBctaJ Coun-j;
Run Tine
CHI u#l cotta H*«n
RMS
20448
10004
19251
8.00 Sec
•2.13i M/S
2,923 M/S
0-j — -J
-15.0
1.7 18.3
Uclocity 1 M/S
Figure 2a. These are typical bar graph outputs from the phase
doppler analyzer. Along with the velocity and size distributions are
mean velocity, mass mean diameter (D30), Sauter mean diameter
(D32), etc. These values were obtained along centerline of spray.
33.00
22.30
10.00
JJ -2.50
S
-13.00+-
1.4
Di*net«ir> uM
Attenpts =
28448
Validations =
10004
Corrected Count =
19251
Run Tine =
8.00 Sec
RrlthMetiC Mtan (D10>= 3.2 uM
«r*JJ M*,in (620)= 4,0 uM
UoluMS Mean (D30>= 5,3 uM
Sluter Mean (&32>= 9.4 uM
Probe = 3.33E-003 om2
NuMber Denstja = 4.19Et0e3 /cc
Uol. = 1.84E-007 cc/s
VoluMe Flux
= 5.53E-0O3 co/s/cm2
CHI O.locitU Mean = -2.131 M/S
- - - - _ 2.923 M/S
Figure 2b. A plot of the mean velocity as a function of droplet
diameter.
85-11
Figure 3a. Plot of mean velocity and rms values as a function of
radial position. These data were taken 5mm from nozzle
Figure 3b. Plots of the Sauter mean diameter (D32) as a function of
radial position.
85-
Water Pressure 90 psig at Inlet
J
Flow Rate(lnncr Ring)
Flow Rate (Outer Ring)
Air Pressure at mlct (psig;
Figure 4. Plot showing the split of air between the inner ring and
the outer ring. This plot indicates the air is evenly split between the
two rings.
1990 USAF-UES SUMMER FACULTY RESEARCH PROGRAM
GRADUATE STUDENT RESEARCH PROGRAM
Sponsored by the
AIR FORCE OFFICE OF SCIENTIFIC RESEARCH
Conducted by the
Universal Energy Systems, Inc.
FINAL REPORT
Pattern Recognition: Machine vs. Man
Prepared by;
Academic Rank:
Department :
University:
Research Location:
USAF Researcher;
Date :
Contract No:
Thomas Abraham
Instructor
Natural Science 6e Mathematics
Saint Paul's College, Virginia
USAF WRDC/AART-2
Wright-Patterson AFB OH 45433-6543
Timothy D. Ross
27 July 90
F49620-88-C-0053
PATTERN RECOGNITION: MACHINE VS. MAN
by
Thomas Abraham
ABSTRACT
Sixty- six images of varying degrees of pattern-ness were compiled and
their AFD* ratings obtained. The respective human ratings were also
gathered from 11 people. The data was consolidated and studied. Also,
through the survey, we investigated to see whether there is a generally
understood meaning for pattern-ness among the respondents.
* Ada Function Decomposition (program)
86-2
ACKNOWLEDGEMENTS
I wish to thank UES for giving me another great opportunity to tap my
research potentials and to enable me to contribute my share in meeting the
objectives of the AFOSR. Also, I would like to thank Les Lawrence and
Paul Johnson for the part they played in realizing this activity.
John Jacobs and his crew of the AART-2 group were very cooperative in
accommodating me and to share the facility with me. Dr Tim Ross deserves
to be thanked for the outlining and the facilitating of the project. He
should also be complimented for his publication which allowed me to
measure pattern-ness of my images mechanically. Dr Tom Gearhart, another
UES Fellow in our section, needs to be thanked for helping me obtain hard
copies of my patterns and for his valuable counsel. Also, Lt. Tim Taylor
is to be thanked for allowing me to use some of his patterns to add to the
variety of my collection. Again, I like to extend my thanks to Mike
Noviskey, Dave Gadd, Keith Graves, Mark Boeke, Michael Chabinye, Shannon
Spittler, and Dr. Mike Breen for expending their time to rate the
patterns. A special thanks goes to Kip Turner and Peggy Alltop for
sharing their facilities with me many a time. Last, but not least I like
to thank my home institution. Saint Paul's College, for the support
rendered.
86-3
I.
INTRODUCTION:
Although this was my second opportunity to participate in this program,
I was still having apprehensions about the kind of work I will be doing,
as my exposure in research was not extensive. However, my meeting with
Dr. Tim Ross during my pre-summer visit enabled me to gather momentum.
My interests in the study of patterns greatly overlapped with those of
Dr. Ross.
The road to success was not quite easy. I had to overcome more than one
impediment to meet ray objective. I launched a sojourn to computer graphics
as I never experimented on one before. With some effort, I managed to
produce patterns on the screen. Getting hard copies of the patterns so
obtained posed another threat to my efforts. Sympathizing in my
predicament. Dr. Tom Gearhart volunteered to help me get the desired hard
copies .
While recognizing the part played by others, I should also attribute my
success to my background education and my strong commitment to fulfilling
my obligations. This laboratory is now involved in Pattern Based Machine
Learning, and I wish to contribute my share in realizing this goal.
864
II.
OBJECTIVES OF THE RESEARCH EFFORT:
Our Objective is to find how closely the machine* ratings of certain
patterns are related to the corresponding human ratings of the same.
Also, we want to investigate the criteria for pattern-ness as evidenced by
people.
* Pattern Based Machine
Published by Tim Ross
Learning (Final Technical Report)
et. al (1990)
86-5
III.
Pattern is a primitive concept experimented by mathematicians,
philosophers, artists, and so forth. Often we think of patterns of
numbers and objects. The very set of integers is achieved through pattern
and order. As one mathematician (ref. 1) has said, "God created integers;
everything else is the work of man". He must have meant that the set of
integers is as patterned as the universe, a creation of God.
Identifying pattern- ness in objects or numbers may be called pattern
recognition. People are good judges of patterns although there is some
degree of latitude between one another. Of course the pattern-ness
changes as the perspective is altered. The more we focus, the less
patterned the object becomes.
Studies on pattern recognition produced considerable revolution in
improving the respective disciplines. However, a general technique that
is rigorous and mathematically oriented is still being experimented upon.
Dr. Tim Ross, in his publication (ref. 2) makes an attempt to accomplish
this by designing an algorithm for pattern recognition and refers it as
'recognizer design'.
Recognizer design is about designing machines that 'map large (possibly
infinite) sets called 'measurements' into small (always finite) sets
86-6
called 'class -labels' , when the dominant factor is the pattern-ness of
this mapping. A mapping, of course, is a set of ordered pairs on a domain
and a codomain such that every element in the domain has a unique paired
element in the codomain as in a functional relationship. A class-label is
also a straight forward concept which means simply the name of the class.
A measurement, however, requires special explanation. It is, in the
conventional sense, an intermediate internal state of the recognition
machine and is considered an important part of the design process.
Measurements are usually reckoned with lenses, microphones, and other
transducers rather than with digital computers. In short, it is simply
the set of recognizer inputs and are not under the designer's control.
This measurement to class -label mapping concept led Dr. Ross to design his
Function Decomposition algorithm.
Recognizer design is partitioned into two sub-problems and termed
'definition' and 'realization'. Definition is used to mean the problems
involving knowledge acquisition, and realization is used to mean those
involving hardware/execution. The generic term 'cost function' is being
employed to reflect the performance and dollar cost considerations for the
entire process including the cost considerations for errors and rejections
due to knowledge acquisition or hardware/execution. Some of these costs
may trade off against each other, and some may not. The net effect of all
these constitute the cost function. Thus the recognizer design problem is
translated in terms of cost function. It is this cost function that I was
able to obtain through the Ada Function Decomposition program.
86-7
IV.
Two dimensional patterns of varying degrees of pattern-ness were generated
by manipulating functions of six or eight variables. (Some sample images
and their respective programs are included in the appendix.) 66 of them
were compiled for the study. Upon obtaining their AFD ratings as a cost
function, people were asked to rate them in terms of the pattern-ness on a
scale 0 through 9. The arithmetic mean of these ratings from 11 people
were found. After standardizing the AFD rating and the people rating to 3
significant figures, they were subjected to our study. (The consolidated
table so obtained is included in the list of tables.) We have noticed
that several pictures had one AFD rating but distinct people ratings. We
took their mean and their standard error and represented them graphically
showing 99.73X confidence level. Also we estimated the correlation
coefficient of the regression line obtained from their scatter plots and
found to be .78 approximately. Again, we classified the people ratings
into classes of width 100 and recorded the corresponding mean AFD ratings
in each class and plotted them against the mid values of the corresponding
classes of human ratings. The correlation coefficient measured .96 here.
(Both graphs are appended in the list of figures.) We also obtained the
scatter plots of all the 66 pairs (graph attached) and their corresponding
regression lines. The correlation coefficient was found to be .80.
86-8
Our experiment was also intended to obtaining a generally understood
meaning for pattern-ness. This was achieved through the questionnaire by
asking the evaluators to comment on their view of the term. (A copy of
the instructions to the evaluators is included in the appendix.)
V.
FINDINGS :
The wide angle between the regression lines is due to the distortion
caused by the gap between the two sets of ratings of low-patterned
pictures. Among the highly patterned ones, there is more consensus
between man and machine. The correlation coefficients we obtained is
fairly high. But for the abnormal cases, we would have obtained a better
figure. The human ratings are intuitive and spontaneous and hence have
considerable variations. The large number of patterns involved also
should have been responsible for the lack of precision. With all these
shortcomings, we noticed a steady trend of direct linear relationship
between the two sets of ratings as the correlation coefficient stayed in
the range .78 through .96.
It was observed from the responses to the questionnaire that the criteria
for pattern-ness centered around symmetry, repetition of the same shape in
an orderly fashion, and easy description. This is in agreement with the
author's perception on pattern-ness: simplicity. According to him,
"having to list all the elements (a brute table look-up) is the essence of
an unpatterred set." Symmetry implies simplicity because you can complete
the whole given its part. The same rule applies to repetition too. Easy
description also implies simplicity gg.9
VI.
RECOMMENDATIONS :
Our experiment is a sure indication that the AFD ratings are quite
reliable. However, there is still room for improvement. A rigorous
technique to measure pattern-ness of any image is still to be explored.
We hope that it will eventually be realized by the same token.
86-10
REFERENCES
1. Kronecker, Leopald (b. Dec. 7, 1823, Liegnitz, Prussia - d. Dec. 29,
1891, Berlin)
2. Pattern Based Machine Learning (Final Technical Report)
by Tim Ross et. al (1990)
86-11
REGRESSION LINES
PATTERN-NESS RANKING EXPERIMENT
Tom Abraham 7/90
PATTERN-NESS RANKING EXPERIMENT
Tom Abraham 7/90
TABLE 1
Pattern-ness Ratin2s of Pictures
Anoroximated to
3 SiE.
FiEures
Picture
Human
AFP
Picture
Human
AFP
Picture
Human
AFP
01
373
625
23
709
766
45
300
625
02
236
000
24
591
453
46
227
000
03
236
453
25
664
813
47
245
438
04
445
000
26
455
000
48
491
719
05
718
812
27
227
000
49
227
000
06
382
641
28
382
625
50
845
812
07
282
375
29
482
438
51
700
656
08
555
438
30
800
812
52
618
781
09
255
000
31
527
625
53
482
625
10
427
000
32
655
797
54
582
797
11
655
797
33
218
375
55
491
625
12
309
375
34
545
792
56
409
375
13
800
812
35
727
812
57
409
484
14
555
438
36
400
438
58
318
375
15
209
000
37
300
422
59
791
922
16
500
656
38
227
000
60
373
375
17
545
781
39
609
828
61
645
781
18
218
000
40
527
672
62
227
000
19
791
812
41
845
812
63
236
000
20
391
375
42
209
000
64
191
000
21
391
438
43
436
625
65
309
375
22
218
000
44
327
438
66
555
828
86-15
APPENDIX MAY BE OBTAINED FROM
THE AUTHOR OR UES
86-16
1990 USAF-UES Summer Faculty Research Program/
Graduate Student Research Program
Sponsored by the
Air Force Office of Scientific Research
Conducted by the
Universal Energy Systems, Inc.
Final Report
Some Results in Pattern- Based Machine Learning
Prepared by:
Academic Rank:
Department and
University:
Research Location:
USAF Researcher:
Date:
Contract No:
Mike Breen, PhD
Assistant Professor
Mathematics Department
Alfred University
WRDC/AART-2
Wright-Patterson Air Force Base
Ohio 45433
Tim Ross, PhD
July 23, 1990
F49620-88-C-0053
Pattern-Based Machine Learning
by
Mike Breen
Abstract
The principle on which fire control hardware and software operate is that of a
function. Our purpose is to find the fastest, most efficient form of the function
possible. It is not feasible to have the function in (input, output)-form. We want to
find the rule under which the function operates. Towards this end, we concentrate
on machine-learning systems and function decomposition. This paper lists several
results in each area seeking to establish a lower bound on training set size at which
the machine-learning system is still performing well. There is no universal lower
bound for each function and each machine-learning system, but we give instances
where a convergence conjecture holds. Also, we give a proof of the main result in
function decompostion.
87-2
Acknowledgements
I am of course extremely grateful to the Air Force Systems Command, the Air
Force Office of Scientific Research, and to Universal Energy Systems, Inc, for this
tremendous opportunity. I found this summer to be a very rewarding experience
professionally.
As individuals, I would like to thank Dr. Tim Ross who gave me all the help and
advice I needed as well as John Jacobs and Paul Johnson who were also very helpful.
Many thanks to Dave Gadd and Devert Wicker who helped overcome the inflexibility
of LaTeX and to Gary Lukens who generated the diagrams of Figure 1. All the people
at AART-2 were very friendly and helpful.
I. INTRODUCTION:
In Avionics, both the amount of information ncecessary to operate effectively, and
the speed with which that information must be processed and acted on to maintain
high performance, require more than sophisticated hardware. What is required is a
design to allow the hardware to perform. An algorithm to allow it to work its best.
The Systems Concept Group of the Mission Avionics Division is studying Pattern-
Based Machine Learning in an attempt to have the machine learn as much as possible
with a small amount of data. The idea is to allow the machine to train on the given
data and find the pattern on this sample. Then the machine can apply that pattern
to all the data. With this approach, a machine can perform its task with less training.
It is the belief of the Group that any function of interest will be patterned. Func¬
tions that most people are fam liar with always have a pattern in that they can be
specified with a rule. In geneial, most functions are not patterned, however most —
if not all — functions that are dealt with in real life are patterned.
Many of the functions discussed in machine-learning are Boolean functions. My
research area is the area of Boolean Matrix Theory — a theory which has many
applications. It can and has been used to analyze dominance relations and food web
relations in the animal kingdom and to analyze cliques and the spread of information
in human social structures. Familiarity with this theory contributed greatly to my
understanding in this assignment.
87-4
II. OBJECTIVES OF THE RESEARCH EFFORT:
When most people speak of pattern theory, they speak of patterns present in
individual objects. For example, they would like a machine to be able to recognize
the letter “A”. Our goal is to bring the idea of pattern to functions (which are
correspondences between objects).
We would like our hardware and software to find the pattern in the assignment
function. The more data the machine has, the better it will do. Yet we do not want
to increase the data a hundred-fold if it causes only a 1/100 per cent increase in
learning. In this regard we would like to know when it is not practical to continue
training. During most of this research period, I worked on this idea. We wanted to
see when a particular inequality was true and when it wasn’t. Both experimental and
theoretical work was necessary in this area.
At times we also worked in the area of function decomposition — also an investi¬
gation into the patternness of a function. We related this idea to the above one since
function decomposition is one example of a machine learning system. We wanted to
provide a different perspective on, and a different proof of, the Basic Decomposition
Condition.
III. THE CONVERGENCE QUESTION
a. In (Ross, 1989) we have the following situation. A function is given which we
want our machine learning system (MLS) to learn. It is impossible to tell the MLS
the whole function. However, we can give it small parts of the function, and allow it
to train on that sample in an effort to learn the whole function. How big a sample
should we give the MLS? This is the focus of this section.
b. We are given a function / : X — > Y, |A| = n. Without loss of generality,
assume Y C R”'. Denote a sample (called a training set) by T where T C X and
|T| = k. Let A = {all functions learned on sets of size k}. The symbols T',k',A',
denote the corresponding notions for larger training sets.
87-5
Define a measure of the difference between two functions as
D{f,g) = E < /(®) - /(®) ~ ^(®) >
X
and define
D(f,A) = '£,D{f,a).
a
Let e = — We want to know at what point, if any, is it true that e can not be
reduced significantly by training on larger sets. In practice, both k and k' are very
small when compared to n.
In (Ross, 1989) it is shown that D(.,A) is minimized by the function a =
That is, the average of the a’s minimizes the total distance. Thus
We call this last number d. Note that d gives us a lower bound on e. We use this to
formulate the convergence question: Given /, is there a k such that for all A'
D{f,A) D{a,A) D{f,A') D(f,A) D{a,A) ,
Ml Ml - Ml ' Ml Ml ■
In this section, we investigate this question. The above inequality is often referred
to as “the inequality”. The inequality determined by the middle term and the sum
on the right is called “the right side of the inequality” with a similar meaning for “the
left side of the inequality.”
c. Write ar for a function in A trained on a set T. We call/ monotone if whenever
%CTCrcX, then D(a;^,,/) < D(ar,/).
Proposition 1 If f is monotone, then for all A and A' .
The proof of this proposition is given by Tom Gearhart. It tells us that if / is
monotone then the right half of the inequality will be true. In fact, there is no need
for the d term.
Although monotonicity is a satisfying idea, we found few sophisticated learning
systems which could be guaranteed to be monotone. So, without monotonicity, the d
term is often needed for the right side of the inequality to hold.
87-6
d. Now let / : X — > {0,1}. In this case we call / a Boolean function. For
® € X define aff(x) to have the value that a majority of the a’s have at x (in case
of a tie, choose 1). Note that an alternate way to find ag(x) is to first find a(») and
then round. To make it easier, I wiU right a for ag from now on, when speaking of
Boolean functions. Define (for fixed x)
Pj = |{a : a(x) = a(x) and a{x) - /(®)}|,Px = Px/\A\-,
Qx = |{a : a(®) = Q!(®) anda{x) ^ f{x)}lq^ = Qxl\A\\
Rx = |{a : a(a:) 7^ ol{x) and a{x) = /(®)}|,rj = Rxl\A\)
S'x = |{a : a{x) a(®) and a{x) 7^ /(®)}|, J* =
and E = {x : f{x) = a(®)},17 = {® : f{x) ^ Q!(®)}. Note that D{j3^A) > D{a,A)
for all Boolean /3; and when ® 6 = 0, when x U^Px = Sx = 0. Primes
appended to these symbols carry the corresponding notions to A\
Define n = - a = r = s =
ueime p — |g| ,q — |y| ,r ~ ,s- |^| .
Proposition 2 If for all A' i) a' = a and ii) s' < 3; and if Hi) r > 1/4 ; then the
inequality holds for all A'.
Proof Since a' = a, E' = E and U' = U. The inequality becomes
\U\{q -r)< \U\q' + W < \U\{q + r) + 2|Pl3
which simplifies to
- p) < W + < |t^l +
The right side is true because of assumption ii) and the fact that g' < 1. As for the
left side, since r > \jA,q — r < 1/2. Because of the definition of a,g' > 1/2. Hence,
the left side holds as well. The proposition is true.
e. Perhaps the first assumption in the last proposition is too harsh. Let us
eliminate that requirement, for indeed the Boolean average may change as the training
set size increases. Now the inequality becomes
\U\{q-r)<\U'\<f^\E'\s'<2\E\s^\U\,
87-7
To focus on the important quantities, let A: = (where (7 7^ 0) and A = (|17| —
\U'\)I\U\. Note that A < 1. Using these terms and the fact that r = 1 — g, we can
write the inequality as
2g — 1^(1 — A)g^ *{■ (Ai H* A)s^ ^ Iks -h 1.
As is often the case with any MLS, the right side of the inequality will be true under
certain reasonable assumptions. We endeavor to show what proportion of the time
the left side holds given that the right side holds.
We assume that both A > 0 and k > 1. That is, we assume that as we take larger
training sets, the Boolean average of these functions will be at least as close to / as
the current Boolean average is. The latter assumption means that |jE| > \U\.
We graph the lines in the q'-s' plane whose equations are (1 — A)q' + (A; + A)^' =
2g — 1 and (1 — A)q' + (A: + A)^' = 2ks + 1. Observe that the inequality is valid in
the region bounded by the lines just mentioned (call them CD and AB, respectively)
and the lines with equations q' = 1/2, g' = l,s' = 0 and s' = 1/2. We determine
the proportion we are interested in by finding the areas of the corresponding regions.
The slope of both CD and AB is Since 0 < A < 1 and A: > 0, the slope of these
lines is nonpositive and for large k, close to 0.
We are interested in the points of intersection of AB and CD with the rectangle.
The s'-coordinate of the point of intersection of CD with g' = 1/2 is
(2,-1)-([1-A]/2) 1 - (|1 - A|/2) l + A ^
«! + A - /c + A - 2(1 + A) ' ■
The s'-coordinate of the point of intersection of AB with g' = 1 is > 0. There¬
fore, the shape of the region under consideration can have one of nine forms. These
forms are given in Figure 1 along with the quantities that determine those forms. The
reader should not be deterred by some of the proportions in the following proposition
as we will estimate them later.
87-8
4q - 3 4q - 3 S A S 2(1 - q) A > 2(1 - q)
y
W
FIGURE 1 I
87-9
1/24- 1'24- 1/2
Proposition Z If A > 0 and k > 1, then
is given below where the numbered proportion corresponds to the region number in
Figure 1.
1. 1 ,
2 1- _ (MJrf _
^ (1-A)(8fes+3A+1) >
f\ Sks+S{l-q)
8fca+3A+l '
4. 1 ,
5 1 _ A+lq-3)^ _
^ 2(fc+A)(l-A)-(fc-'l»-A)2 >
/? 1 _ (l-A)(3A+8g-7)
^ 2(fc+A)(l-A)-(it-lfcs-A)2 '
7. 1,
o 1 _ A+‘i<?— 3)^
^ 2(fc+A)(l-A) '
Q "ik-A+T-Hn
2(fc+A) •
Proof Proportions 1, 4, and 7 are obvious. Observe that in the remaining regions,
the s'-coordinate of C is . In all other regions where B is pictured, its
s'-coordinate is ■ In regions 2 and 3, the s'-cooidinate of A is — In
regions 5 and 6, A’s g'-coordinate is . For D, when it is the qf'-intercept
(regions 2, 5, and 8) its ^'-coordinate is In regions 3, 6, and 9, the s'-coordinate
ofDis^.
The areas of the regions follow from these facts.
We now get a lower bound on the more cumbersome of these terms. In region 2,
A < 2(1 - 9) so,
(A4-4g-3)^ ^ _ (1-A)^ ^ 8fes-b4A
(1 — A)(8A!s 4- 3A + 1) (1 ~ A)(8fcs -b 3A -b 1) Bks -b 3A -b 1
In region 5, the proportion is larger than 1 - = 2fe+3'A-i
k - Aks < 1). In region 6, again k — Aks < 1, so the proportion given is at least as
large as Using the same fact as we did with region 5, we find
that the proportion corresponding to region 8 is at least
Summarizing this, we have
87-10
Proposition 4 If A > 0 and k> 1 then the 'proportions in the numbered regions are
at least
8ita+3A+l Vi
(region 5),
(region 6), and
2A:+3A-1
2(A:+A)
(region 8).
We believe that k should be the dominant term in these expressions. These
results are intended to be a guide as to when much improvement is not likely (when
the proportion is high), thus making further training undesirable,
f. Because of the wide range of possibilities for f and for the MLS, the inequality
may or may not be true. In this section and the next we maintain the assumption
that / is Boolean and look at two different MLS’s. For the first MLS, the inequality
holds. For the second, it doesn’t.
Define the first MLS as follows. For a; € Jf, if x € T, aT(x) = /(®); if ® ^
X,P{ar{x) = f{x)) = j. That is, we have an MLS which assigns function values to
the learned functions in a random, independent way for elements not in the training
set. We call this a random MLS.
Proposition 5 Given a random MLS with the property that P(ax(x) = /(®)) =
P(aT(x) -f- f{x)) = 1/2 for x ^ T, then P{f — a) > 1 — 1/n whenever {k -
2)900*/^“'* < n < 10''^ (where k > Z) or more simply k > IZ and 2/: < n <
Proof By the definition of a, the probability that a(x) = f{x) is the probabihty that
|{a G A : a{x) = /(a;)}| > |{a G A : a{x) 7^ /(®)}|> Thus our desired probability is
the probability that the number of functions which are trained at x plus the number
of functions which are not trained at x but which are equal to f{x) is greater than
the number of functions which are not trained at x and which are not equal to /(»).
There are C(n,k) functions in A, C(n-l,k-l) of them are trained at x. Therefore,
P{a{x) — f{x)) is the probability that more than of the C(n-l,k) functions
87-11
not trained at x are equal to f{x). Let F represent the random variable ’vhose value is
equal to the number of functions in A not trained at x which are equal to f{x). Note
that the values of F are binomially distributed with mean C(n-l,k)/2 and standard
deviation \JC{n — l,fc)/2. Since p = 1/2, we can get a good approximation to the
desired probability by using a normal distribution. The usual adjustment of the
F- values by 1/2 is negligible in this case.
So, P(a(®) = f{x)) «
If n > {k — 2)900‘'''(*'”^^ + 1, then using rough estimates, we get the above to be
at least as large as P{z > -30) which is at least 1 — 2.78 x 10“®® (Smirnov, 1965).
Hence, we can be fairly certain that given x,a(x) = /(x). If fc > 13 and n > 2k, then
n > 2(k — 2) + 1 > (k — 2)900*/^*^"®) + 1 and our conclusion is satisfied.
Because of independence, P(a = /) = (P(a(x) = /(®))]" > (1 - 2.78 x 10“®®]".
We can use a Taylor polynomial (for g(y) = y") to get an approximation. Let
e = 2.78 X 10“®®. Then (1 - e)" « 1 - nc, with an error of at most . By
assumption, n < 10 '^, so the error term is at most 3 x 10“®® and
F(a = /)>!- 2.78 x 10“'' - 3 x 10“®® > 1 - 10“’'^ > 1 -
n
If we have a random MLS for which P{a'f{x) = f{x)) >1/2 when x ^ T, then
the mean number of learned functions equal to f{x) will increase, and the standard
deviation will decrease. So P{a = /) will be larger. This gives us the following
corollary.
Corollary 6 If a random MLS has the property that for x ^ T,P{a'r{x) = f{x)) >
1/2, then with n as in proposition 5, P{a = /) > 1 — 1.
So if the function the MLS is attempting to learn is Boolean, and if the MLS is
random doing at least as well as chance, then for almost all domains and training
sets, we can be fairly certain that a and / are equal.
z >
i!-2A:(n-n! _ C(n-l,k) \
2(n-fc)!fc! 2
^/C(n-\,k)
87-12
Once we have a domain size and training set sizes which satisfy the conditions of
proposition 5, then U' = U = %. Therefore, the inequality reduces to 0 < s' < 2s.
When we neglect the d term on the right in the original inequality we get the stronger
statement 0 < s' < s. We now set about to find when these inequalities are true,
naturally concentrating on the right side of each.
Proposition 7 If the MLS is random and P{aT(x) = f{x)) = 1/2 for x ^ T,
then P(s' < s) « 1 for each A' where k < k' < 2n and n satisifies the conditions of
proposition 5.
Proof If A: = &', the conclusion follows.
Otherwise, fi,> = So, n,,.,
n~k'
2 _
n^k
Ps' -
n-k'
k-k'
2n
-k
Also, a:
2 _
s) = P(s' - S < 0) >
Hence, ^ and P(s' <
<
V
k'-k
_2a_
n-k'
+
n-k
= P\z<
k'-k
n^C(n,k') ^ n2C(n.fc)
LEEl
V C(n,
k'
k>)
I n—k
C(n,k) ,
(the strict greater- than sign arises because we are calculating a smaller area than
the one we would calculate usually). Since n > 2k and A; > 13 > 5, < 1/32.
Because k < k' < nl2,C{n,k) < C{n,k') and the entire denominator which contains
these two terms is at most 1/4. Thus, P(s' < s) > P(z < > P(z < 4) « 1.
If for each individual x, the probability that each trained function agrees with
f(x) when x is not in the function’s training set is more than 1/2, then the z-score
above would increase, and the stronger version of the inequality would again be true.
Corollary 8 fVith a random MLS for which P{af{x) = f{x)) > 1/2, we have that
P{s' < s) ^ 1 for each A! where k < k' < nl2 and n satisfies the condition in
proposition 5.
We summarize these results below.
87-13
Theorem 9 Suppose
.{0,1},
2. our MLS is random,
3. for each x ^ T and for each ar, P{af{x) — f{x)) > 1/2,
4. |T| > 13 and 2k<n< IQ-^^;
then
Djf.A) _ D{a,A) ^ D{f,A') D{f,A)
Ml Ml - M'l ~ Ml
for all A’ where jT'l < n/2.
Corollary 10 The inequality in the convergence question is true under the same
assumptions that appear in the theorem.
Proof Obvious.
g. Let us now look at another MLS for a Boolean function, Let mhc the function
in A such that D(m, /) < D{a, f) for all o € A. That is, m is the closest function in
A to / — sometimes called the last best guess. Let our learning system learn in this
manner. Given T' C X and a' € A',
a'{x) —
This MLS is monotone by definition, hence the right side of the inequality holds
because of proposition 1. The left side is more difficult — it need not hold. First
note that, at a given x, calculating the real average of the a'-values at x, gives us
. C(n - l.k' - l)f{x) + [C{n, k') — C{n — l,k' - l)j7n(s)
a ( X ) = - - — ~T - =
C{n, k')
_ Since we assume that n > 2k', we see that a'(x) = m(x) for all x.
So for X t f/, <5jr = |{a € A : X ^ Ta}\ = C{n- l,k) (where Ta means the training set
/(x) if X € r
Tn(x) if X ^ T'
87-14
that a is trained on), = \{a £ A : z £ T„}| = C[n — 1, A: — 1). For all »,5x = 0.
The left side of the inequality becomes
yr C{n — 1, A:) - C{n - 1,A! — 1) ^ C{n — IjA:')
C{n,k) - C{n,k')
which is equivalent to |f7|(n — 2A!) < \U'\(n - k’).
If our last training was done on sets of size k,\U'\ = |17| - k' (assuming the last
best guess misses f by more than k' points). The inequality we are interested in
becomes \U\{n - 2k) < (|C/| — k')(n — k') or k'(n — k') < jf7|(2As — k'). We regard k
and k' as being of the same magnitude, so that the n — k' term is dominant and this
inequality fails. If we train on all sets of size A: + l,fc + 2,...,A;', then \U'\ will be even
smaller and again the left side of the inequality fails.
IV. FUNCTION DECOMPOSITION
a. To connect this area with the last, we now make note of the fact that we wrote
a Turbo Pascal program which acts on our function decomposition database to gather
information about the truth of the convergence question for that learning system. It
is just now being run. The program calculates both the real average and the Boolean
average, compensating for any undefined points and finds d for each average. Dr.
Tim Ross wrote most of this program.
We also point out that of all the MLS’s we investigated, the ADA function de¬
composition algorithm is the top performer when /: N — > {0,1}. We now look at
function decomposition.
b. Our work here was done first in the mechanics of the ADA function decom¬
position algorithm. In order to learn a function (/), this algorithm chooses the least
complex function which agrees with the original function on the training set. We
wanted to know the expected size of the set of functions which were both no more
complex than / and agreed with / on the training set.
If |A’’| = n, and G = {g\g : X — > {Ojl}}) then |(j| = 2^^"^ If the training
set T C X has size k and I ■= {g ^ G g(x) = f{x) for all a: € T}, then |/| =
For a randomly chosen subset of G,H,P{H contains exactly r elements of
87-15
I) = 1 . The random variable r has a hypergeometric distribution.
Therefore E{ir\H)='^. In our case, E{I H) =
c. Our idea of patternness of a function is based on its complexity cost — the
lower the cost, the higher the patternness. The cost is related to how the function can
be broken into simpler pieces. That is, how it decomposes. The principal result here
is the Basic Decomposition Condition. To prove it in the most general case, we must
define a relation on the columns of the function’s partition matrix. The definition of
that relation follows.
To modify the semantics of the problem, we define a total function f : Xi X x
. . . X Xn Y U{u} such that /(a!i,®2>---»®n) = • • • >®r>) when / is defined
and f{x\ , X2, . . . , Xn) = “ when / is undefined.
A “column” for some fixed Vx and V2 is defined as the sequence: Cviv^ =
(/(vi,V2,6(l)),/(vi,V2,6(2)),/(t;i,U2,6(3)),...,/(vi,V2,6(|V3|))) where 6 is a bijec-
tion from V3 into {1, 2, . . . , IF3I}. Columns form a set of columns for a fixed
V'ifSvi = {Cuiujlui € Vi}. When / is total, the column multiplicity (t/) is the maxi¬
mum over V2 in 14 of {[5v2)}; however, when / is not total, we need an extra step.
Call two columns compatible if the only coordinates in which they differ are those
where either is undefined. Consider Sv^ = {C'„,„j|/(vi,V2,V3) is defined for each
i;i,i;3}. If 5„j is empty, we can go to later steps. If not, define on S„, using
prefix notation by if f(v,,V2,V:i) = f(v,>,V2,V3) for all V3 G V3.
This is an equivalence relation on Enumerate the resulting equivalence classes
E\ , E2, . . , E,,^^ calling representatives ej, 62, . . . , . Also enumerate the elements
of the set 5„j\5,,5 : C| , C2, . . . , Choose the first class such that Ci is compatible
with the representative of that class and adjoin C\ to that class. If no such class is
found, C\ will belong to its own class Ep„^+\. If Ci creates a new class, it is the
representative - called Otherwise, define the new representative of the class
that Cl is in, to be e[ where (e[)n, is the value of (CJm or (efc),„ if either or both
are defined (if both, they must be equal), and {e|.)„, is undefined if both (C'l),,, and
(ek)m are undefined. Here, m — I to ll3i and m stands for the “coordinate” of the
87-16
column vectors. Representatives of other classes remain the same but are denoted
individually as e^. Continue in this manner with the other elements of S„^\Svj. That
is, if a column Ca is not compatible with any of the existing representatives, it will
create a new class whose number is one more than the number of the last class:
Z„j + 1. It will be the representative, denoted e^^j. If a column Ca is compatible
with an existing representative, we choose the first occurrence of this, adjoin Ca to
that class and proceed as with Ci except in the above steps (when we dealt with
Cl) replace ek with with and ej with ef. Finally, given V2, we have the
set of classes Evj = < hi £ which partitions the set of columns. The
classes’ representatives are Call the equivalence relation
determined by this partition of 5„j , Mvi •
Now we define = |Fu.J. When is empty, there is only one since can
not occur in the expression. Finally, we define v as the maximum over all va in V2 of
i/„j . This definition relies only on elementary set theory for background.
d. Now that the relation is defined we use it to prove the following theorem. The
theorem is not new, but the proof is.
Theorem 11 The Basic Decomposition Condition: Let f : X\ >: X2 x - ••>< Xn — * F
be a partial function where each X, is finite. Let Fi,F2,h:}, be a partition of the
domain. There exists a set V and functions (j) : V\ x V2 — ^ Z and F \ Z 'KV2 x
such that when f is defined f{vi,V2,vTl = F(<^(v|,?;2)) ^2, V3) if and only if v J jZf.
Proof First show that 1/ < \Z\ implies the desired existence.
Step 1. Define 0 : Fi x -» 2 by <i>{vx,V2) = i if Cu,„j €
Step 2. Define F : Z XV2 xV:^ —*Y us follows:
i. if z = i for some i, then F(i,V2,U3) = (^i"pb(v]) where b is the bijection from V3
into {1,2,...,|V^j1} ,
ii. otherwise,F(z,U2,U3) is undefined.
Once in place, defined coordinates of the representatives of the equivalence classes
do not change, so when defined /(vi,V2,W3) = = F(<f)(vi,V2),V2,V3).
87-17
Now we prove that the existence of functions / : X Vj — > Z and F : ZXV2XV3
Y such that whenever /(wi,V2> ^3) is defined, f{vi,V2,V3) = -^(/(ui, ^2)? implies
that u < \Z\. First observe that v < is logically equivalent io <\Z\ for all
V2 € ^2- Assume to the contrary that there exists V2 € V2 such that i/y, >|Z|.
1) We can assume that equivalent columns correspond to ordered pairs which
have the same inverse image under (f>. When (7t,„5 and are iZu^-related, then
when defined f{v\,V2,V3) = /(vi<,t;2,t;3) for each V3. Thus, F{(i>{vi,V2),V2,V3) =
F(^(vi», V2), ^2,^3). No harm is done to the relationship between f and F and the
range of (j) is no larger than before. Hence, we assume that the inverse image of an
element of Z contains ordered pairs which correspond to an equivalence class (under
) of columns.
2) If t/yj > |Z|, then there must be two non-equivalent columns whose corre¬
sponding ordered pairs have the same image under <f). These columns can not come
from 5uj since Rv^ is the equality relation there because if two columns from 8,,^ are
not equivalent, there must exist a V3 such that f{vi,V2,V3) = /(vi',V2, V3). Hence,
f{vi,V2,V3) = F{<f>{vi,V2),V2,V3) = V2), V2, V3) = /(«!', ^2,^3) 7^ /(Vi,V2,V3),
a contradiction.
3) For 5uj\5„j, use the ordering developed in defining i?„j. Choose the first element
of Svi\Sv2,Cj, such that the number of equivalence classes in U {Ci , . . . , (7j} is
larger than U {Ci, . . . , Cj) )|. If no such element exists, i/„j < \Z\. Because of
how Cj was chosen, Cj must create its own •''“.ss. That is, Cj is not compatible with
any representative of the equivalence classes at that time. We can write Cvivi for Cj
and see that there exists U (C,, . . . , (7j} such that (^(^1,^2) = (f>{vii,V2).
Call the representative of the class of We know that is not compatible
with e“. Therefore, there is a Cv^„v■i currently in the class with Cv^,v^ such that for
some U3,/(vi'', ^2,^3) and f{vt,V2,V3) are defined and unequal. This last statement
follows from the fact that a column is compatible with a representative of a class if and
only if it is compatible with each element in the class. So we have Rv^{Cv^,|V2^
which implies that ^(ui'',U2) = and that F’(^(ui»,U2), V2) ^3) =
87-18
F{<f>{v^,V2)iV2,Vi) for each v-^. Yet for some V3, f{vi'i,V2^V3) ^ f{vi,V2,V3) and both
are defined. As before, this contradicts the assumption about / and F. Hence, no
such non-equivalent columns exist. The theorem is proven.
V. RECOMMENDATIONS:
a. The convergence question is a very broad one. We have seen that it can not
always be answered in the affirmative. Because machine-learning is an imprecise area
now, no ground rules are given for an MLS. In order to be able to frame the question
in a more specific way, such ground rules are needed. It may just be a matter of time
before such rules are given.
b. We have given some results on machine-learning when / is a Boolean function.
Some complicated functions can be reduced to the Boolean case, but others can’t.
Investigation into other types of functions is a good idea.
c. Function decomposition is our criterion for patternness. How do other experts
in the field feel about this criterion? Some effort should be given to sounding out those
experts whether by publishing our results and waiting for feedback, or by attending
conferences and meetings. An encouraging sign that function decomposition is a
valid criterion for p?tternness is the behavior of the ADA function decomposition
algorithm (FDA) in the area of machine-learning. It may be ‘he paramount practical
justificiation for the group’s definition of a patterned function.
d. Why does FD.4 work so well? Much more sophisticated MLS’s (neural nets,
for example) are outperformed by FDA. Research should be done into the reasons
behind the success of FDA. Such research could unlock secrets in both areas: machine
learning and function decomposition.
87-19
REFERENCES
1. Ross, Timothy D. “A Convergence Method for Evaluating Whether or not the
Training Set for a Machine Learning System Contains an Appropriate Number of
Samples”. December, 1989.
2. Smirnov, N.V., Tables of the Normal Probability Integral, the Normal Density
and its Normalized Derivatives, New York, The Macmillan Company, 1965.
1990 USAF-UES SUMMER FACULTY RESEARCH PROGRAM/
GRADUATE STUDENT RESEARCH PROGRAM
Sponsored by the
AIR FORCE OFFICE OF SCIENTIFIC RESEARCH
Conducted by the
Universal Energy Systems, Inc.
FINAL REPORT
PROBABILISTIC IR EVIDENCE ACCUMULATION
Prepared by:
Dr. R. H. Cofer and Jim Perry
Academic Rank:
Associate Professor and Research Assistant
Department and
Electrical Engineering and Computer Engineering
University:
Florida Institute of Technology
Research Location:
WPAFB, WRDC/AARA, ATR Branch
USAF Researcher:
Jim Leonard
Date:
September 22, 1990
Contract No:
F49620-88-C-0053
Probabilistic IR Evidence Accumulation
by
Dr. R. H. Gofer and Jim Perry
ABSTRACT
The work reported here is a continued outgrowth of Bayesian Target Recognition research
started in the 1989 Summer Faculty Research Program. During the current 1990 research
effort, the emphasis has been on probabilistic evidence accumulation in the IR ATR prob¬
lem. Two important and fundamential types of probabilities were found: underlying target
temperatures, and spatial homogeneity of target temperatures. The first is important for
target to decoy discrimination, while the second can overcome unavoidable lack of the tar¬
get’s thermal history. Correctly used, these two probabilities will result in overall consis¬
tency of IR evidence accumulation. Also shown is the general robustness of probabilistic
evidence accumulation to practical considerations of uncertainty, ignorance, and function¬
al approximation.
88-2
Acknowledgments
We certainly thank the Automated Target Recognition Branch of the Wright Research and
Development Center at the Wright-Patterson Air Force Base and the Air Force Office of
Scientific Research for sponsoring this research. Universal Energy Systems should also be
mentioned for their efficient administration of this contract.
We would like to thank Ed Zelnio, Jim Leonard, Lori Westerkamp, and Bill Foley for al¬
lowing us to participate as members of their research team. Their numerous discussions
and guidances provided the format for this reseaivh. Kevin Willey is to be sincerely
thanked for his assistance in integrating the SUN workstation which we brought with us
into the AARA computational environment and keeping it working during a condnuing se¬
ries of lab changes. The Wright Research and Development Library is to be commended
on its excellent current store of technical research material.
88-3
1.0 INTRODUCTION
The central problem in ATR research is how to accumulate modeling and imagery evi¬
dence in a consistent and meaningful form. Earlier research by Dr. Gofer has abstractly
showed that a probabilistic approach is both consistent and capable of the highest recogni¬
tion performance - provided that the required conditional probabilities can be found.
Two types of probabilities are required in a complete probabilistic evidence accumulation
process:
• The probability that a specific region in the image could have arisen from the target, and
• The probability that the same region could have arisen from non target background.
The first type of probability can be immediately utilized in a wide variety of ATR designs
and is thus of more immediate interest to AARA. Last summer’s research showed method¬
ology of how to obtain this probability for LADAR imagery. This summer’s research
shows methodology of how to obtain this probability for IR imagery, even under igno¬
rance of the target’s thermal environment.
2.0 OBJECTIVES OF THE RESEARCH EFFORT
For Dr Cofer, the goals of this research effort were to: extend the theory of probabilistic
evidence accumulation, develop probabilistic modeling approaches to the IR target prob¬
lem, and supervise the efforts of his research assistant, Jim Perry. For Jim Perry, the goals
of this research effort were to: convert an AARA IR modeling program written for a per¬
sonal computer to the considerably higher performance UNDC / C / Sun View environment,
learn the rudiments of IR target modeling in preparation for Master’s research, and to par¬
ticipate in an advanced research team.
3.0 IR MODELING AND EVIDENCE ACCUMULATION
As planned, the majority of the summer’s research effort went into the IR target modeling
and evidence accumulation arena. An IR target imagery base was placed on the SUN
88-4
workstation for further consideration throughout the upcoming academic year. Dr. Gofer
and Jim Perry greatly expanded their knowledge of IR target modeling in preparation for
future continued interaction with Lori Westerkamp and Bill Foley. Jim Perry successfully
converted a non-trivial IR modeling problem from the BASIC IBM AT/EGA environment
to the much more powerful C SUN SPARC workstation Sun\tiew environment for a 20:1
improvement in speed, graphic clarity, and expandable problem size. Dr. Cofer successful¬
ly determined the fundamentals of splitting the probabilistic IR evidence accumulation
problem along a thermal history / temperature homogeneity cleavage. The research meth¬
ods used and results obtained are further discussed below.
Jim Perry’s coding is documented in the final program documentation left at AARA. Dr.
Gofer’s results are further documented in a 40 viewgraph Probabilistic Evidence Accumu¬
lation package also left at AARA.
3.1 Physical IR Target Modeling
The physics of IR target modeling. Figure 1, has been repeatedly investigated over the
years, but not with the view of developing ATR evidence accumulation probabilities. Thus
several points are reasonable to note as they have a bearing on the development of the re¬
quired probabilities. The IR image is formed as a result of seeing the temperatures, T’s, of
the target. Much if not most of the information of the target’s IR appearance derives from
time varying heat flows, qi’s, to and from the target’s environment. If these flows were to
go to zero, the target would become constant temperatured throughout and thus detail
within its silhouette would be difficult to detect. Fortunately, heat flows within and outside
the target are non-deterministic making the target’s surface temperatures visible but, un¬
fortunately, hard to analyze. This non-deterministic behavior then is the genesis of the
need to place a probabilistic flavor to the modeling of the IR target, its temperatures, and
heat flows.
Physically, IR modeling inevitably begins with the equation
88-5
(EQl)
dt ^^2
which continuously relates target temperatures and heat flows across space and time. Eq. 1
£r _ 3^ Y^qj
Figure 1. The IR Physical Modeling Problem
thus leads directly to the resulting spatial temporal model of Figure 1 lower left. Although
such a model is quite general, it provides few easily accessible handles with which to at¬
tack the probabilistic IR problem.
3.2 The BIOT Condition in IR imagery
One of the best methods of obtaining clues to the nature of data is to examine it with an
open mind. While looking at a chronological IR image history of a stationary M60 tank,
88-6
Dr. Gofer became struck by the fact that he could see the tank wheels, hubs, webs, tracks
and body. Even more striking was the fact that various tank components were appearing
and disappearing from visibility as whole entities. These observations lead to further study
of heat transfer in solids. This study resurfaced the BIOT condition - objects (or subob¬
jects) which have a high internal to external heat flow rates can be essentially considered
constant temperatured across their mass. More interestingly, BIOT conditioning was quite
evidently also being seen in the IR image. Wheels, webs, tracks and body of the tank all
had high internal conductivity relative to their external heat flows. Thus each had relative¬
ly constant although non-deterministic temperatures across their mass. When their non-de-
terministic temperatures were sufficiently diflferent, parts could be distinguished one with
respect to another. When the non-deterministic temperatures were not sufficiently differ¬
ent, the parts remained visually indistinguishable.
Recognizing that BIOT isolated parts exist within a target allows the model of lower left
Figure 1 to be reduced to the more spatially discrete model at the lower right of Figure 1.
Although investigators have often discretized target thermal models, this current research
specifically notes that the BIOT discretization also shows up in the IR imagery.
3.3 Target Reduction to an Equivalent Circuit
Further study of past IR modeling efforts revealed that the thermal heat transfer properties
of metallic solids, on either a macro discrete or micro infinitesimal level, can be modeled
as an electrical circuit composed of resistors and capacitors. The resistors serve as flow
paths, the capacitors serve as heat sinks within the solid, electrical currents into and out of
the circuit serve as environmental heat flows to and from the solid and finally electrical
currents within the circuit serve as heat flows within the solid. As a result, the model of
lower right Figure 1 and upper right Figure 2 can be replaced by a resistor capacitor cir¬
cuit, represented graphically in the upper right of Figure 2. As a result the pixels of the im¬
age from an IR sensor are equivalent to voltage readings taken from points in a resistor
88-7
capacitor circuit being driven by environmental heat flows - the qj’s at the top of the cir¬
cuit^ of Figure 2.
SPATULLY DISCRETE
TEMPORIALLYCONT.
SUBMODELS
WE ARE SEEING A SNAPSHOT
ANALOGOUS TO A SINGLE
VOLTAGE PROBE OF POINTS
IN AN ELECTRICAL CIRCUIT
BEING DRIVEN BY
UNSEEN NONLINEAR
FORCING FUNCTIONS!
Figure 2. An Equivalent Model of What Is Being Sensed
3.4 Probabilistic ATR Requirements
To understand the precise nature of the probabilistic characterization needed in the IR ATR
problem, consider Figure 3. Upon receipt of the incoming IR image, R]r, the recognizer
must determine the probabilities, P(Rir 1 1, Hx) and P(Rir 1 1, Hy), of Rir conditioned on
1. This problem should not be confused with a similar problem that has long been addressed in electronics.
In electronics, one is interested in the time characterization of the (usually single) output of a circuit which is
driven by a (usually stationary ergodic) input. In the IR modeling problem, we do not have multiple time
looks at a single output. Rather we have a single time look at multiple outputs where the inputs are driven by
nonstationary, nonlinear nonergodic driving functions. This is a very different type of characterization prob¬
lem.
88-8
indexing information, I, and target type hypotheses, and Hy. To do this, the recognizer
RECOGNITION IN THE IR
TARGET DETECTION PROBLEM
MEANS "—mini, III t
^WEATHER, SOLAR LOADING. CLOUD COVER
' TEMPERATURES OF SURROUNDINGS. ETC
■^^'^Iiere.the.dri^te'toye
l>^h'6h*det|niiiiiisdeToitinK'^ y
LIMITED DIRECT nnsiTPVATTnisK^
IMAGE FORMiVnOCi _
MODEL ' ■; "v ■ ^
TARGET X I'
hiL-j'' x'-'TsLi'
IMAGE FORMATION
MODEL
-J'., . target Y
P(Rir1I.H^
P(RirII,Hy)
image \
\ IMAGE RESULTS
formatioM—^ recognizer
Figure 3. The ATR Scenario
needs the functional forms, P( RirI I, H^) and P( Rq^ 1 1, Hy), of the two probabilities,
P(Rir 1 1, Hx) and P(Rir 1 1, Hy). Given P( R^r ! I, H^) and P( 1 1, Hy), the recognizer
v.'ill then insert Rip^ to obtain the two scalar values PCRjs 1 1, H^) and P(Rir 1 1, Hy) of in¬
terest to the final decision. Thus the forms, P( Rjr ! I, H^) and P( R|]^ ! I, Hy), are the prob¬
abilistic characterizations the ATR needs from the IR modeling effort.
3.5 Probabilistic IR Modeiing
The problem which now arises is “How to determine in detail the fonns, P( Rjr 1 1, H^)
and P( RipJ I, Hy)?”
88-9
Figure 3 shows the conditional probabilities, P( Rjr 1 1, Hx) and P( Rir 1 1, Hy), arc each
simultaneously:
• a deterministic function of the modeling circuit of the respective targets, and
• a non-deterministic function of the thermal environment within which the target is im¬
mersed.
The second point is crucial. Although the non-deterministic thermal environment acts di¬
rectly on the image, Rir, received by the recognizer, the modeling formation process will
have very limited observation of the thermal environment. Basically the reason is that in a
battlefield engagement, it will be impossible to obtain a time history set of observations of
important thermal environmental parameters such as solar loading, wind and target mo¬
tion. Thus the solution must be probabilistic.
3.5.1 An Early IR Probabilistic Problem Decomposition
88-10
Figure 4 shows the first approach that was taken toward calculation of IR modeling proba¬
bilities. It was based upon the concept that the physical target model can be reasonably
“chunked”, due to the BIOT condition, to allow the probability statement to be written in
the conditional form shown. The concept, as then envisioned, was that consideration of
smaller target components could be somewhat isolated from major thermal environmental
p(R I Tj.g^ wheel ’ "^front wheel » ^ , I, H)
Figure 5. Transforming The Problem By Switching Sample Spaces
effects by conditioning with respect to the larger target body. While a step in the correct di¬
rection of breaking down the probability calculation in terms of more elemental quantities,
this formulation could not explicitly break out the effects of the weather and climatic ther¬
mal environmental influences. As a result the detailed modeling determination of P( Rjr I
I, Hx) and P( Rir 1 1, Hy) would still remain a Herculean task.
88-11
3.5.2 A More Satisfactory IR Probabilistic Problem Breakdown
After considerable further effort, the much more satisfactory probability breakdown of
Figure 5 was found. The main utility of this breakdown is a separation of the probabilistic
effects of space and time in the thermal target behavior dictated by Eq. 1.
By conditioning on the underlying temperatures of the object, P(Rir 1 1, H), for each H,
can be conditionally partitioned. The first element of the partition is a probability PCTj^af
wheel- Tfront wheel- ^body 1 1- H) depends only upon the surroundmg thermal environ¬
ment. The other elements of the conditional partition are probabilities, P(Rrear wheel ^ ^rear
wheel- H)- P(^ffont wheel ^ ^front wheel- H), P(Rbody ^ ^body ^ ^)- depend only on
the underlying temperatures of their own individual target body components. Thus the
probabilistic calculation is cleanly split into distinct subproblems: the underlying tempera¬
ture and the homogeneity problems. Further, Figure 5 points out that the modeling circuit
can also be split into individual components for separate simplified consideration. In each,
the underlying temperature of the part is introduced as a voltage forcing function.
The joint probability of the underlying target temperatures do not depend in any way on
such non-deterministic variables as sensor noise, atmospheric effects, and emissivity vari-
a/ions. This of course is a further indication of the clean split. The determination of the
joint probability of underlying target temperatures is taken up in the next section.
The set of IR observational probabilities conditioned on the underlying target tempera¬
tures^ do not depend upon the surrounding thermal environment or each other^ - giving
further evidence of the clean split of the original problem. Their determination is further
considered in Section 3.7.
2. In the example of Figure 5, these were P(R„^ *hcei ' wheel. I H), PfR^on, wheel 1 Tf,o„t wheel- 1 H) and P(Rbody i
Tbodyll. H).
3 . It .s realized that there are other usually second order mutual couplings between target body components, chiefly radi-
aLon. Often such interaction will be small and even if not. further research should be able to extend the degree of decou¬
pling.
88-12
3.6 The Underlying Part Temperature Subproblem
In summary form, Figure 6 shows the research work that went on toward the characteriza-
R H/ Lori - Developed System Simulation Concept
* Lori/ Ed • Located Correlations in M60 Thermal History Imagery
Jim Perry / Bill • Converted IK line program to C / Sun
* Bili / R H • Developed Math Model of M60 Tank Wheel
’"Ongoing Theoretical Considerations
Too large a driving space? Law of large Numbers? Gaussian?
Stochastic Processes? Can the answv i- ever be found? Does the answer
even need to be found?
Figure ti. Consideration of The Underlying Temperatures of a Target
tion of the underlying joint target temperature probability. Starting with a system level
concept, there was a desire to develop a corresponding simulation as an aid and cueing
mechanism for further thought. Key to this is the state variable graph of Figure 6. If a real
(or simulated) non-deterministic thermal time environment is input to the deterministic
target thermal model as shown in the accompanying system block diagram to the left, one
will obtain the representative target temperatures Tj through Tn - also as a function of
time. These target temperatures can then be parametrically plotted as shown by the coiled
solid line of the graph. Each coiled loop represents the joint underlying temperatures of a
The fact that the loops, or equivalently joint daily temperature cycles, do not deterministi¬
cally repeat are due to incidental daily non-deterministic variations of cloud cover, season,
etc. This was confirmed by work done by Lori Westerkamp. Working from real M60 tank
IR imagery data taken over a period of time, she plotted actual diurnal cycles which did
not, in fact, repeat.
As the state variable temperatures track out their daily variations, they repeatedly cross
any specific time of day that an image might be taken, as represented by *’s on the graph.
At any one time of day, sufficient numbers of these points will begin to define a joint prob¬
ability density function of the underlying target temperatures - such as the PCTjear
Tfront wheel* Tbody* Figure 6. Note that the resulting joint temperatures are not just
dependent upon conditions at the time of the image, but also upon conditions for many
hours past. Obviously this dependance slowly dies out as one goes backwards in time.
Note then that the underlying diurnal temperatures of the target are really just filtered ob¬
servations of the limit cycle of the strange attractor of Nature’s thermal cycling‘s. This
gives hope that the joint probability density function of the underlying target temperatures
can be reasonably stated in terms of fractal theory.
A first prototype of a graphical interface was developed during this summer to allow du’ect
simulation and visualization of the diurnal state variable cycling. Provisions were made:
• to allow changes to environmental parameters,
• to allow seeing the sweep of the cyclic behavior from one to another limit cycle, and
• to gather up time points visually showing the development of the underlying target joint
pdf.
4. Initial reviews of fractal theory were made in the 1989 Summer FollowOn Research effort on the convic¬
tion that fractal theory plays a role in the domain of ATRs.
88-14
Due to the shortness of time, the capability was demonstrated using a different fractal than
the thermal environment/target model - since the thermal environment/target model was
still being simultaneously developed.
The Academic/AARA research team made many strides to provide reasonable thermal tar¬
get models during the summer. Key factors here were speed of computation combined
with reasonable accuracy since a large number of daily cycles must be computed before
one begin to see the emerging pdfs. Jim Perry working with Bill Foley, an AARA IR ex¬
pert, took one of Mr. Foley’s IR computer models written in BASIC for the IBM AT color
graphics environment and converted it for use in the much higher performance SUN work¬
station environment. R H Gofer working with Mr. Foley helped to develop a BIOT condi¬
tioned heat transfer model of a lower wheel assembly of the M60 tank from physical
measurements in the field. Mr. Foley constructed the first draft of the model in BASIC
computer code. Dr. Cofer took this data, reconverted it back to a theoretical form and sug¬
gested several changes to improve certain aspects of its accuracy. After further verifica¬
tion, he then modified it for traceability of heat flows, and was in the process of developing
detailed documentation at the end of the research period. The next two steps which should
be taken are to code the model within the SUN environment and to develop means of sim¬
ulating a chronological thermal environment of months in duration, most likely along frac¬
tal lines. Upon completion of these two steps, an actual simulation can be run that will
begin to give a indication of the nature of the joint pdf of the underlying target tempera¬
tures. This indication can then be used as a cueing mechanism for guiding analytic re¬
search toward a more mathematical derivation of the joint pdf.
Figure 6 also shows several of the ongoing positive and negative theoretical considerations
that are currently considered likely to play to a role in future heat transfer simulations. Of
these the theory of large numbers is most likely to provide a means by which the pdf can
be analytically found. Once major periodicities of the diurnal cycle have been removed, it
is also hypothesized that there will remain large numbers of relatively independent non-
88-15
deterministic variables. These will be summed by the thermal filtering properties of the
target to give, upon proper probabilistic conditioning, a limiting pdf - perhaps even with
Gaussian tendencies.
Figure 6 then poses the question, “Can the answer ever be found?” Obviously the joint
pdfs, if obtainable, will be of use and particularly so in the case of target-decoy discrimina¬
tion where geometries can be the same but thermal properties will differ. On the other
hand, it is certain that the underlying target part temperatures are functions of a seemingly
endless list of time and space dependent non-deterministic variables: cloud cover, shadow¬
ing, rain, wind, target orientation, target position, terrain, mud plastering, variations of
paint emissivity, internal heating, target manufacturing variabilities, etc. It will not only be
a potentially difficult problem to characterize these non-deterministic driving functions,
but in a realistic battlefield situation one may not even know the full list. Additionally, it
can be challenging to develop multi-dimensional pdfs due to the sheer size of the mathe¬
matical spaces. In summary, the fundamental limitation of developing joint target temper¬
ature pdfs is ultimately dependent upon the nature of the thermal environment. This nature
is still in question and unverified. If it can be characterized sufficiently well, along fractal
or other lines of thought, then joint target temperature pdfs may be indeed be obtainable.
The deeper question posed by Figure 6 is “Does the answer (to the underlying target part
temperature question) even need to be found?” Dr. Gofer’s continued inspection of actual
IR imagery and related study indicates that the answer may very well lie in the area dis¬
cussed in Section 3.8.
3,7 The Homogeneity Subproblem
The homogeneity probabilistic characterization subproblem is defined as determining the
form of the probability of the image regions of target components conditioned on underly¬
ing target component temperatures. In Figure 5, these probabilities are P(Rrear wheel ^ 7’rear
wheel’ H), P(Rfront wheel ^ T'front wheel’ ^(^body ^ T'tody ^ H).
In the case where the BIOT condition reigns and emissivity holds constant across the sur¬
face, homogeneity probability calculations will be easily developable as a function of the
sensor image chain. For other less common cases where BIOT conditioning does not hold
and/or emissivity does not remain constant across the surface, further research will be re¬
quired. A portion of the summer was devoted to amassing the necessary research materials
to continue such an investigation though the ensuing academic year.
3.8 Robustness Given Lack of Knowledge of Underlying Temperatures
Jim Leonard, as responsive and responsible AARA technical contact for this research ef¬
fort, was particularly adroit at setting one squarely on the horns of the dilemma: “How do
you expect to complete the probabilistic characterization of IR targets, given the manifest
problems in determining underlying target temperatures?” A fair question to which much
thought was given and the following answer was developed: “Estimate the underlying
temperatures from the image data itself and then use this estimate to developing the final
probabilities, P(RirII,H), for each H.”
Figure 7 portrays the idea. Figure 7 (a) shows the normal breakdown of P(RirII,H) and in¬
dividual sources of information. Figure 7 (b) shows an alternate breakdown where tlie un¬
derlying temperatures are taken from estimates on the image itself. Two questions can be
raised of this alternate breakdown:
♦ “What is the effect of errors in the functional form of the probabilities?”, and
• “What is the error introduced by the approximation of the underlying temperatures
themselves.”
Research during the period indicates tiiat increasing the image resolution can overcome
both concerns. The obvious question is “How much must the image resolution be raised?”
By example, an answer of not much. Shown in Figure 8 is a feature extraction problem: “Is
the image region a rear wheel of a truck or background?” The domain knowledge and rec-
88-17
ognition setup information is also shown. The wheel temperature, T^, is estimated by its
mean taken from the image. The resulting recognition probability, P^y, is plotted as a fam-
the homogeneity the underlying temper-
subproblem ature subproblem
(a) Normal Probabilistic Breakdown
IMAGE ^
MODEL
P (R,r1 I, H) = ll? (Ri (i) I Ti, I, H) ^i| I, H)x [[? (Rj (j) [ T^, I, H) P(T2l I, H)
(b) Maximum Likelihood Probabilistic Breakdown
Figure 7. Alternate Homogeneity Subproblem Breakdowns
ily over a range of probabilities of misiecognition of background, Pf. The abscissa is a nor-
/
malized function of the number of pixels on the wheel, N; the true difference of
temperature of wheel from background, M; and the standard deviation of the sensor
IMAGE
noise, a. The solid line shows the accuracy if the true underlying temperature were per¬
fectly known. The dotted line gives the recognition performance when the temperature is
estimated from the image itself. Obviously, the drop of performance is small and is cor¬
rectable by a very small increase in image resolution. Note no modeling of underlying
temperature probabilities was required, tiius that problem was avoided.
The above example is representative of the robustness generally achievable by estimating
the underlying target temperatures from the image itself. Of course, research should be
continued to further extend the above results.
88-19
4.0 RECOMMENDATIONS FOR FURTHER RESEARCH
Probabilistic evidence accumulation is a most promising methodology for use in ATR de¬
sign and implementation. It has withstood each theoretical recognition performance chal¬
lenge so far in a consistent manner and has provided guidance for multisensor fusion,
handling lack in target thermal history, allowing meaningful incorporation of modeling in¬
formation, etc. Probabilistic evidence accumulation should also prove particularly useful
in coping with the strong speckle/mutual interference probabilistic component of SAR im¬
agery.
The Academic/AARA research team should continue its joint investigations into:
• Computer performance modeling and simulation of LADAR, IR and LADAR/ER target
recognition based upon target and image chain modeling,
• In depth real world consideration of the variability of IR target skin effects including
emissivity, thermal conductivity, and directional radiation characteristics under varying
conditions of rain, humidity, splattered mud, cloud cover, etc.,
• The probabilistic effects of optical blur,
• A spatial probabilistic temperature representation of metallic surfaces under nearly
BIOT conditions, and
• The probabilistic representation of background clutter.
88-20
1990 USAF-UES SUMMER FACULTY RESEARCH PROGRAM
GRADUATE STUDENT RESEARCH PROGRAM
Sponsored by the
AIR FORCE OFFICE OP SCIENTIFIC RESEARCH
Conducted by the
Universal Energy Systems, Inc.
FINAL PEP.0RT
Inyggtlqationg .of ,a Lov^r Boun<a
on thg .Error in Lgarng<a-Fungti<?ng
Prepared by;
Academic Rank:
Department and
University;
Research Location:
USAF Researcher;
Date:
Thomas K. Gearhart, Ph.D,
Associate Professor
Department of Mathematics, Computer
Science, and Physics
Capital University
Avionics Laboratory
WRDC/AART-2
Wright-Patterson AFB OH 45433-6543
Timothy D. Ross, Ph.D.
27 July 90
Contract No;
F49620-88-C-0053
Investigations of a Lower Bound
on the - Error, in Learned Functions.
by V
Thomas K. Gearhart
ABSTRACT
An exact expression for the difference between the
average sum*?*pf-squares error for a collection of learned
functions and a lower bound oh that error is derived. A
bound bn the difference of the average sum-of -squares
errors for two distinct collections of learned functions is
obtained-. This bound can be computed without knowledge of
the desired function outside the training sets. A condition
is isolated which assures that average sum-ofrsquares error
will decrease as the size of the underlying training sets
increases. The lower bound on the average sum-of-squares
error is experimentally compared with the traditional
measure of error for specific machine learning systems.
89-2
I wish to thank tt® Air Force Systems Command and the
Air Force Office of Scientific Research, whose sponsorship
provided me this opportunity to participate in a
stiroulating research project. I also thank the staff of
the SFRP Office at Universal Energy Systems for their
careful and helpful handling of the administrative aspects
of the program.
Numerous individuals in the Avionics Laboratory at
Wright-Patterson AFB contributed to making my stay
enjoyable and productive. John Jacobs and his staff in the
System Concepts Group were most hospitable and provided a
very pleasant work environment. I am especially grateful
to Dr. Tim Ross for his assistance in identifying an
interesting, and challenging research problem, and for
providing, direction to my investigations throughout the
Summer. Conversations with Mike Breen and Lt. Tim Taylor
were helpful in clarifying certain issues related to my
project. The assistance of Shannon Spittler and Peggy
Alltop in analyzing experimental data and preparing graphs
is also gratefully acknowledged.
I also wish to thank Dr. Harold Brockman, Chairman
of the Department of Mathematics at Capital University, for
his encouragement and support of my application to this
program ,
89-3
I. INTRODUCTION
Since January, 1989,. Dr. Timothy Ross and associates
in the Avionics Laboratory have been engaged in an
extensive study of Pattern Based Machine Learning (PBML) .
A central finding of the study is the existenqe of a strong
connection between pattern-ness and the concept of function
decoropositipn. Techniques and tools from both mathematics
and computer science have played an important role
throughout this research effort.
My initial technical training was in pure mathematics
with a concentration in general topology. More recently, I
completed eleven months of graduate studies in computer
science during a sabbatical leave granted by my university.
These studies^ included courses in digital logic, pattern
recognition, and automata theory. Each of these three
fields has a direct connection to some aspect of the PBML
study. My recent research interests in computer graphics
are also complementary to the PBML study from the
standpoint of developing programs for the visualization of
the phenomena under study.
89-4
II. OBJECTIVES OF THE RESEARCH EFFORT
Machine learning systems (MLS) attempt to predict the
global behavior of a function using ihfomation about the
function on a small sample of its domain. An important
issue is the determination of how large such a sample
should be. In [4], Dr. Ross describes a method for
determining a lower bound on the average sum-of-squares
error resulting from samples of a given size. During a
meeting prior to the start of my research period, it was
decided that I would develop ideas introduced in this paper
and attempt to resolve various conjectures advanced
therein. My approaches to meeting these objectives
involved the application of mathematical techniques and the
development of computer programs for the simulation of
various learning systems. These approaches are discussed
in sections III and IV.
III. THEORETICAL INVESTIGATIONS
In this section mathematical techniques are applied to
develop ideas introduced in [4].
Let f;X -> R^ be a function from a finite set X into a
Euclidean space. The notation f|rp will be used to denote
the restriction of f to subset T of X. A Machine Learning
System ( MLS ) is a system which uses a restricted function
f|ip = { ( X, f(x) ) 1 X € T } to generate a function
aip:X -> R^. A function aip so generated by an MLS is called
a learned function associated with the training set T. The
89-5
perforaahce of an MLS can be measured by how closely such
learned functions approximate the function f over the
entire domain X.
To make more precise the notion of closeness of
functions, define the sum-of-squares difference between
functions f and g as:
d(f,g) = S II f(x) - g(x) ||2
xeX
where the double bars indicate the usual Euclidean norm.
If A is a collection of functions from X into R*', define
the sum-of-squares difference between function f and a set
of functions A as:
d(f,A) « E d(f,a)
aeA
Let |A| denote the cardinality of A. The following theorem
establishes an important relationship between d(f ,A) and
d(a,A) where a is the average of the functions in A.
Theorem 1 Let X be a finite set and suppose f:X -> R^.
Let A be a finite collection of functions from X into r”.
Define the function a by a(x) = ( S a(x) ) / jA].
aeA
Then d(f,A) - d(a,A) = |A| d(f,a).
Proof In this proof the notation <V,W> will denote the
usual inner product of vectors V and W in R^.
d(f,A) - d(a,A)
= S S ||f (x)-a(x) ||2 - S S ||a:(x)-a(x) ||2
aeA xeX aeA xeX
89-6
h
= S E
aeA X6X
= S S
X6X aeA
*22
xeX aeA
( l|f(x)-a(x)|j2 - ||a(x)-a(x)l|,2 )
( <f(x)-a(x) ,f (x)-a(x)> - <a(x)-a(x) ,a(x)-a(x)> )
(
*22
xeX aeA
<f(x),f(x)> - 2<a(x),f(x)> + <a(x),a(x)>
- <a(x),a(x)> + 2<a(x),a(x)> - <a(x),a(x)> )
( <f(x),f(x) - 2<a(x),f(x)>
- <a(x),a(x)> + 2<a(x),a(x)> )
Since the inner product is distributive over vector
addition and 2 a(x) = |A|a(x), the above double sum
aeA
reduces to:
* 2 (
xeX
|A|<f(x),f(x)> - 2|A|<ot(x),f(x)>
- I A| <a(x) ,flt(x)> + 2|A|<a(x) ,a(x)> )
= |A| 2 ( <f(x),f(x)> - 2<f(x),a(x)> + <a(x),Qt(x)> )
xeX
= [a] 2 <f (x)-a(x) ,f (x)-a(x)>
xeX
= [Al 2 |lf(x)-a(x)|l2 = |A| d(f,a). |
xeX
Corollary 1 d(f,A)/|A| - d(a,A)/|A| = d(f,a) .
Proof Divide both sides of the equality in the conclusion
of the theorem by |a|. |
Corollary 2 d(f,A)/|A| >d(a,A)/|A|.
Proof The right side of the equality in corollary 1 is
nonnegative . |
Corollary 2 above has important consequences in the
context of machine learning systems. Suppose the MLS is
89-7
attempting to learn a function f . Let A be the collection
of learned functions output by the MLS when trained on a
collection of training sets of a given size. Then the
corolla^ guarantees that d(a,A)/|A| is a lower bound on
the average sum-of-squares error d(f/A)/|A|.
Since in practical applications of machine learning
systems the function f is known only on the training sets,
d(f,A)/|A| is not computable. However, since each member
of A and the average a is known over the entire domain,
d(a,A)/|A| can be computed. Thus we have a computable
lower bound on the average sum-of-squares error.
In many instances the functions under study are
Boolean functions. A Boolean function f:X t> Y has
domain X» {0,1}*' = (0,1) x (0,1) x ... x {0,1} and
codomain Y - {0,1}. Under these circumstances the
pointwise average a of a collection A of functions might
take on values other than 0 and 1 and hence might not be a
Boolean function. However, a Boolean function a’ can be
defined from a by setting a'(x} = 1 when a(x) > 1/2
and setting a'(x} =0 when a(x) < 1/2. Note that a’(x)
= 1 when a(x) =1 for at least half of the functions in the
collection A and o'(x) = 0 otherwise. The question of the
existence of an analogue of Theorem 1 for Boolean functions
is answered affirmatively by the following.
Theorem 2 Let X be a finite set and suppose f:X -> {0,1}.
Let A be a finite collection of functions mapping X into
89-8
{0,1}. Define the function a;X -> [0,1] by
a(x) = ( S a(x) ) / | Al .
aeA
and define a';X -> {0,1} by a’(x) =1 if a(x) > 1/2
and a'(x) =0 if a(x) < 1/2.
Then 0 < d(f,A) - d(a‘,A) < |A| d(f,a'),
Proof Define subsets S, T and U of X as follows:
S = { X € X I f (X) * a' (X) }
T = { X 6 X I f(x) = 0 and a'(x) = 1 }
U = { X e X I f(x) » 1 and a'(x) =0 }
Note that S, T, and U are disjoint and their union is X.
By definition, d(f,A) - d(a',A)
= S S ( f(x)-a(x) )2 - E S ( cit'(x)-a(x) )2
aeA xeX aeA xeX
= S S [ ( f(x)-a(x) )2 - ( a«(x)-a(x) )2 ] (1)
aeA xeX
Since the bracketed expression in (1) is zero on S,
(1) can be rewritten as:
S S [ ( f(x)-a(x) )2 - ( flt'(x)-a(x) )2 ]
aeA xeT
+ S S [ ( f(x)-a(x) )2 - ( a'(x)-a(x) )2 ] (2)
aeA xeU
Using the definitions of T and U, (2) is the same as:
E S [ ( -a(x) )2 - ( 1 - a(x) )2 ]
aeA xeT
+ E S [ ( 1 - a(x) )2 - ( .-a(x) )2 ]
aeA xeU
= E E ( 2 a(x) -1)+ 2 E(l-2 a(x) ) (3)
aeA xeT aeA xeU
Since S a(x) = |a| a(x), (3) is equal to:
aeA
89-9
|A| ( S ( 2 a(x) - 1 ) + S (1^2 a(x) ) ) (4)
xeT x€U
For X 6 T, 1/2 < a(x) <1 and so,
0 < 2q!(x)-1<1 for X € T.
For X e U, 0 < a(x) < 1/2 and so,
0 < 1 - 2 a(x) < 1 for x e U.
Thus it follows that (4) is bounded below by 0 and
bounded above by |a| ( |T| + |u| ). Finally, note that
since f and a ' take on only values from the set {0,1},
d(f,a') = r ( f(x) - a'(x) is simply the cardinality
xeX
of the subset of X on which f and a’ disagree. But this
cardinality is simply |t( + |u|. Since d(f,A) - d(a',A)
is equal to (4), the theorem follows. |
Suppose d(f,A)/|A| is the average sum-of -squares error
associated with a set A of functions generated by an MLS
training on sets of a given fixed size. Let d(f,B)/|B| be
the average sum-of-squares error associated with set B of
functions generated by the MLS training on sets of larger
fixed size. Dr. Ross, in [4], conjectured that it might be
possible to obtain a bound on ( d(f,A)/|Al - d(f,B)/|B| j
whicn involved d(a,A)/|A|. The following theorem is a step
in this direction.
Theorem 3 Suppose f:X -> r”* where X is a finite set and
||f(x)|| < M for every x e X. Let A and B be two
collections of functions mapping X into and let a and
0 be the averages associated with A and B respectively.
89-10
That is, let a(x) = ( S a(x) ) / |A| and let iS(x) =
( S b(x) ) / tB|.
b£B
Then 1 d(f,A)/lA| - d(f,B)/|Bi | <
2M S ||a(x)-/3(x)l| + ) S (lla(x) (x) IP) |
X6X X6X
+ I d(a,A)/|Ai - d(i3,B)/|B| |.
Proof Using Corollary 1 of Theorem 1 and the triangle
inequality, | d(f,A)/|A| - d(f,B)/|B| |
« I ( d(a,A)/|A| + d(f,a) ) - ( d(/3,B)/lB| + d(f,)3) ) |
< |d(f,o)-d(f,/9)| + I d(a,A)/|A| -d(i9,B)/|Bl ].
Now d(f,a)-d(f,i9) - r ||f(x)-ft(x)||2 - s l|f (x)-)3(x) P
X£X X£X
= S ( <f(x)-a(x) ,f (x)-a(x)> - <f (x)-i8(x) ,f (x)-/3(x)> )
X£X
= S ( <f(x),f(x)> - 2<f(x),a(x)> + <a(x),a(x)>
X£X
- <f(X),f(X)> + 2<f(X),iS(X)> - <PW,P(X)> )
= S ( -2 <f(x) ,a(x)-)3(x)> + ||a(x)||2 - ||)0(x)||2 )
X£X
= -2 S <f (X) ,a(x)-^(x)> + S ( ||a(x)||2 - ||^(x)|l2 ).
X€X X£X
Thus, I d(f,a) - d(f,/3) |
< 2 S I < f(x),a(x)-^(x) > I + I S (l|a(x)||2-||^(x)||2) |
XfiX X£X
By the Cauchy-Schwarz inequality, | < f (x) ,a(x)-)3(x) > |
< ||f(x)|| ||a(x)-/3(x) II < M II a(x) - )0(x) || and
the theorem follows. |
A desirable feature of an MLS would be that d(f,A)/|A|
decrease as the size of the underlying training sets
89-11
increases. The following discussion and theorem establish
a condition under which this will occur.
Suppose X is a finite set and f:X -> R® and let S be a
sample set of size n selected from X. Let a^:X -> R®
denote the function output by a given MLS when it is
trained using tS s. Let us agree to call an MLS monotone
with respect to f and S if d(f,au) < d(f,a(j) for any
two training sets T and U such that S 2 U 2 T. For each
k = 1, 2,...,n let Ajj. * { arp I T is a k-member subset of S } .
Note that IAj^I » C(n,k) where C(n,k) denotes the number of
k~member subsets of an n-member set.
Theorem 4 If an MLS is monotone with respect to function
f:X -> R® and an n-member sample set S 5 X, then
d(f,Aj^)/|A]^| < d(f,Aj)/|Aj| whenever k > j.
Proof Let U be a k-raember subset of S. Since the MLS is
monotone, d(f,au) < d(f,atii) for any j -member subset T
of U. Since there are C(k,j) such j-member subsets of U,
C(k,j) d(f,au) < S d(f,aT).
T c U
lTl= j
Summing each side of this inequality over all k-member
subsets of the sample set S yields;
C(k,j) S d(f,au) < S ( S d(f,aT) )
UCS UCS TCU
|U|= k |U|= k iT|= j
In the double sum on the right, a given d(f,aip) appears
multiple times. This is because a given j-member set T is
contained in several distinct k-member sets U. The number
of occurrences of d(f,aiji) in the double sum is C(n-j,k-j),
89-12
because C(n“j,k-j) is the number of ways of enlarging a
fixed j -member subset of S to a ktraember subset of S. So
the double sum can be replaced by C(n-j,k-j) S d(f,aut)
T c S
|T|» j
Hence, C(k,j) S d(f,au) < C(n-j,k-j) S d(f,aip).
U c S T o S
|U|= k |T|- j
That is, C(k,j) d(f,A}5) < C(n-j,k-j) d(f,Aj).
Since |A]^| = C(n,k) and |Aj| = C(n,j) the above inequality
is equivalent to:
C(k,j) C(n,k) d(f,Ak)/|Ak| <C(n-j,k-j) C(n,j) d(f,Aj)/|Aj|
Since C(k,j) C(n,k) = C(n-j,k-j) C(n,j), the proof is
complete. |
IV. EXPERIMENTAL . INVESTIgATlPI^g
To assist in assessing the practical value of the
lower bound bn the average sum-of-squares error, a computer
simulation of an MLS appropriate for learning real valued
functions was developed. Given a function f and a training
set, this MLS generated an approximation of f by linearly
interpolating the points of the training set.
The simulation was run for several real valued
functions f defined on the domain X = { 1, 2, ..., 1000 ).
In each case a sample set S of size 70 was randomly chosen
from the domain X. Next, for each of the values k = 10,
20, ... , 60, ten training sets of size k were randomly
chosen from S. For each k, the graphs of the ten
functions, the average a, and the target function f were
all displayed on the same axis. Also the following
measures of error were computed:
e »
d »
& »
z
aeA
Z
xeX
|if (x)-a(x) II 2
|A| 1
x|.
Z
aeA
Z
xeX
||o(x)-a(x)||2
|A| 1
|x|
Z
_aeA_
Z ||f(x)-a(x)||2
X€S\T,
|A| (
Ts|-k )
where is the
training set in¬
ducing a.
The measure e is a normalized version of the average
sum-of-squares error and d is the theoretically guaranteed
lower bound on e. The measure S is a measure of error
which has traditionally been used as a figure-of -merit for
machine learning systems.
For this MLS based on linear interpolation, results of
the simulations suggest that % tends to be a better
estimate of e than d is.
A second MLS appropriate for use with Boolean
functions was also designed and implemented in software.
This MLS incorporated a three-layer neural network. For an
n-variable Boolean function, the input layer had n nodes,
the middle layer had 2n+l nodes, and the output layer had a
single node. Weights for the input layer nodes were
permanently fixed at 1. Weights for the middle and output
layer nodes were initialized randomly and then updated
using the backtracking strategy described in [3].
89-14
The effectiveness of the network as an MLS varied
greatly with the function being realized. It did well on
functions with many vacuous variables. For example, on a
seven variable xl XOR x2 function, the error rate was only
two percent with training sets of size 40 from the domain
of size 128. At the other extreme, the network did no
better than chance attempting to learn a seven variable
parity function.
Studies of e, d, and d were also made for the neural
network MLS. The only trend that could be discerned was
that e was frequently between d and e.
V. REggMMEUDAllQJiS
A. If the lower bound on the average sum-of-squares
error is large, then the machine learning system is
performing poorly and/or the training sets are not large
enough. However the converse is not true, and so small
values of the lower bound need to be interpreted with
caution. They need not imply that the actual error is
small.
B. Simulations for the machine learning system based
on linear interpolation indicate that the traditional
estimate of error is often closer to the actual error than
is the lower bound on the error.
C. Comparison of the three measures of error (see
section IV) for an existing machine learning system based
on function decomposition might be helpful -in ciarifying,
the usefulness of the lower bound bn the average
sum-of-squares error.
Di The neural network described in section IV appears
to find some feature of the function other than
decomposability. For a, wide range of functions, the
network was outperformed by an existing MLS based on
function decomposition. A systematic study of other
network topologies and activation functions might reveal
ways of improving the network’s perfomnance as an MLS.
A hybrid MLS using a combination of neural network ideas
and ideas from function decomposition theory might also be
worthy of research.
89-16
REFERENCES
1. Caudill, Maureen, "Neural Networks Primer., Part I,”
AI Expert; bece^er 1987, Vol. 2, No. 12, pp; 46-52 .
2. Caudill, Maureen, "Neural Networks Primer, Part II,”
Ai Expert. February 1988, Vol. 3, No. 2, pp; 55-61.
3. Caudill, Maureen, "Neural Networks Primer, Part III,"
AI Expert, June 1988, Vol. 3, No. 6, pp. 53-59.
4. Ross, Timothy D. , "A Convergence Based Lower Bound oh
the Error in Learned Functions," December, 1989.
89-17
1990 USAF-UES SUMMER FACULTY RESEARCH PROGRAM/
GRADUATE STUDENT RESEARCH PROGRAM
Sponsored by the
AIR FORCE OFFICE OF SCIENTIFIC RESEARCH
Conducted by the
Universal Energy Systems, Inc.
FINAL REPORT
Machine Learning Applied to High Range Resolution Radar Returns
Prepared by:
Lawrence 0. Hall and Steve G. Romaniuk
Academic Rank:
Assistant Professor and Research Assistant
Department and
Computer Science and Engineering
University:
University of South Florida
Research Location:
WPAFB, WRDC/AARA, ATR Branch
USAF Researcher:
Jim Leonard
Date:
August 6, 1990
Contract No:
F49620-88-C-0053
Machine Learning Applied to High Range Resolution Radar Returns
t>y
Lawrence 0. Hall and Steve Gi Romaniuk
ABSTRACT
This report examines the use of a neural network learning algorithm and a hybrid
neural network, symbolic learning algorithm on the problem of recognizing airplanes
from high resolution radar returns. Quickprop and SC-net are the techniques used.
The intent of the study is to determine how to both recognize the planes and recover
the aspect angle in an algorithm with small set up and good space/time character¬
istics. Three different representations of the radar returns to the learning algorithm
were tried. The problem of representation is very important in this study. The
first two representations were geometric hashing schemes. The last is a binning and
averaging scheme. It has shown some invariance to aspect angle shifts, which is
important in limiting the number of training times and examples. In both learning
systems the third representation has been used to get 100% recognition for sp’’ *e
sets of aspect angles.
90-2
Acknowledgements
We would like to thank the Automated Target Branch of the Wright Research
and Development Center at the Wright- Patterson Air Force Base and the Air Force
Office of Scientific Research for sponsoring this research.
We would like to thank Rick Mitchell, Ed Zelnio and Jim Leonard for the valuable
discussions and guidance they have provided during the course of tnis research.
Especially, thanks to Rick for getting the hashing algorithms going. We would like
to thank Kevin Willey for his assistance in getting our computing environment set
up. The access to the data both real and generated and the stimulating problem
itself are much appreciated.
90-3
I Introduction
The returns from High Range Resolution Radar provide data that may be used to
identify target airplanes. Specifically, air to air recognition is an object of this study.
The angle of the target is also of interest. Hence, an examination of how machine
learning can be used to provide recognition of the targets from this type of data is
underway. This report will discuss the issues involved and progress to date.
One important problem is to understand the nature of the data and develop an
appropriate representation for machine learning. A second problem of equal import
is what type of machine learning algorithm will work the best. These two issues are
discussed in the following.
Either a supervised learning or unsupervised learning scheme may be applied.
Supervised learning works well, when the classes can be labeled beforehand and there
is someone with domain information able to guide the learning process. When there
are no labels for classes and/or there is no one available with a good understanding
of the domain area in which learning is to be applied, unsupervised learning may be
the best choice. The issues are more involved, but in this case the domain fits the
criteria for supervised learning schemes.
At the current time there is tremendous interest in neural network learning al¬
gorithms. Neural networks (also called connectionist networks) are an attempt to
model the function of the human brain. When it comes to learning algorithms, some
are very closely tied to the way that the brain is thought to work and others are
only loosely based on a biological model. Neural network techniques for learning
compete with symbolic techniques, such as ID3 [8]. Recent studies have shown that
for some domains performance can be almost identical [10].
Connectionist networks are said to have an advantage in that they require less
set up. However, whether this advantage exists is debated. A connectionist network
can be used without modification in different non-symbolic domains. The appear
90-4
to work well with data that is visual in nature, and patterns. The domain we are
working in provides patterns that may be visually recognized by people (although
this is not currently done in air to air encounters). There is no readily apparent
symbolic information in the domain of this problem. Hence, connectionist models
are a good choice for learning to recognize the planes from the radar returns.
Two different approaches to recognition are being used. One is a hybrid connec¬
tionist, symbolic system developed at tlSF, called SC-net. The other is Quickprop a
feedforward network architecture which back propagates error and changes weights
to minimize the error. Quickprop may be viewed as a fast version of the well known
back-propagation algorithm. The learning algorithms are further discussed in the
following sections.
The data used in this study was generated by the Srcrcs tool developed by the
Syracuse Research Corp. The data used simulates a wide band or high range resolu¬
tion radar (HRR) with a 20 meter range window, consisting of 256 discrete points.
The magnitude of the return at each point corresponds to the radar cross section
at a specific range on the aircraft. There are seven typical airplanes available for
training and testing, the F4, F14, F15, F16, F18, T38 and Lear. The elevation and
aspect angle may be chosen by the user when generating the return. In this study
we have been working with no elevation, only changes in aspect angle.
II Objectives of the Research Effort
The goals of this research study are described in this brief section. The first is to be
able to distinguish among airplanes based upon high resolution radar returns using
some machine learning algorithm which requires minimal set up. Tied up with that
objective is the necessity of determining an appropriate representation of the data
for the learning algorithm.
A second objective is to determine the aspect angle of the plane that is in the
radar return. Another goal is to have training times and the storage required within
reasonable ranges. A further goal is to determine what factors are important in
the recognition process and gain some understanding of how itds being done by the
algorithm.
Ill SC-net
This section briefly describes a hybrid connectionist, symbolic approach to learn¬
ing the type of information that could be used in rule-based expert systems or
sub-symbolic information. The system does its learning from examples. They are
encoded in much the same way that examples to connectionist systems would be
presented. The exceptions are due to our variable representation. The system can
learn concepts where imprecision is involved. The network representation allows for
variables in the form of attribute, value pairs to be used. Relational comparators are
supported. Symbolic rules can be generated from the learned network configuration.
Learning may be simply described by-the following. It uses a network structure,
which is configured based on the distinct examples presented to the system. That
is, the network structure is grown based on the distinctly new examples seen by the
system. Initially, only input and output cells need to be specified. For examples
which resemble others previously seen, bias values of cells in the network are ad¬
justed. The system can learn incrementally. A complete description of learning in
SC-net can be found in [1]. Rules may also be directly encoded in the network [9].
The system will explicitly indicate any contradictions, and patterns that have
already been seen, which has proven useful in this research.
a SC-net - A Fuzzy Connectionist Expert System
A connectionist model is a network, which in its simplest format has no feedback
loops. It consists of three types of cells (input, output, and hidden cells). Every
cell has a bias associated with it, which lies on the real number scale. Cells are
90-6
connected through links which have weights associated with them. In the SC-net
model of a connectionist network, each cell can take on an activation value within
the range [0..1]. This corresponds to the fuzzy membership values of fuzzy sets. The
uncertainty handling constructs come from fuzzy set theory.
In fuzzy logic one may define disjunction (fOR) as the maximum operation,
conjunction (fAND) as the minimum operation and complement (fNOT) as strong
negation. Since fOR and fAND are defined as maximum and minimum operations,
we let certain cells act as max and min functions, in order to provide for the above
operators. In order to be able to distinguish cells as modeling the min (fAND) or
the max (fOR) function we use the sign of the bias of a cell to determine which of
the two functions is to be modeled. Furthermore, we denote a bias value of zero to
indicate when a cell should operate as an inverter (fNOT).
b The Network Structure
We can think of every cell in a network accommodating n inputs /„ with associated
weights CWn. Every cell contains a bias value, which indicates what type of fuzzy
function a cell models, and its absolute value represents the rule-range. Every cell
Ci with a cell activation of CA,- (except for input cells) computes its new cell ac¬
tivation according to the formula given in Figure 1. If cell C; (with CA,) and cell
Cj (with CAj ) are connected then the weight of the connecting link is given as
CWijf otherwise CW{,j = 0. Note, an activation value outside the given range is
truncated. An activation of 0 indicates no presence, 0.5 indicates unknown and 1
indicates true. In the initial topology, an extra layer of two cells (denoted as the
positive and the negative cell) is placed before every output cell. These two cells
will be collecting information for (positive cell) and against the presence of a con¬
clusion (negative cell). These collecting cells are connected to every output cell, and
every concluding intermediate cell (the.se are cells defined by the user in the SC-net
program specification). The final cell activation for the concluding cell is given as:
CAi - cell activation for cell Ci,CAi in [0..1].
CWij - weight for connection betw^n cell C,- and Cj, CWij in R.
GBi - cell bias for cell C7,-, CB{ in [-1..+1].
CA': = I
minj=o,„,i-i,{+i,„n{CAy* CWi,j) * \CBi\ GBi < 0
maxj=Q„„i-i,i+i,„n(GAj * GWi,j) * \GBi\ GBi > 0
1 - (GAj * GWi,j) GBi = 0 and GWi,j 0
Figure 1: Cell activation formula
CAoutput=CApositive_cell+CAnegative_cell-0.5 .
Note, the use of the cell labeled UK (unknown cell) in Figure 2. This cell always
propagates a fixed activation of 0.5 and, therefore, acts on the positive and the
negative cells as a threshold.
The positive cell will only propagate an activation >= 0.5, whereas the negative
cell will propagate an activation of <= 0.5. Whenever there is a contradiction in the
derivation of a conclusion, this fact will be represented in a final cell activation close
to 0.5. For example, if CApositive_cell=0.9 and CAnegative_cell=0.1, then CAout-
put=0.5, which means it is unknown. If either CApositive.cell or CAnegativo_cell is
equal to 0.5, then CAoutput will be equal to the others cell activation (indicating
that no contradiction is present).
IV Quickprop
Quickprop [2] was developed by Fahlman at CMU. It is a back-propagation neural
network architecture. Hence, it is a feedforward network with a learning scheme that
back-propagates the error and then minimizes the error by modifying the weights.
We have been using a 3-layer (one hidden layer) network representation. The system
has been shown to be up to an order of magnitude faster than standard back-
propagation [3]. In other tests that we have run, it has shown itself to also be
90-8
significantly faster than back-propagation with coniparable accuracy after learning
[2], Quickprop has been the connectionist network of choice in this study due to its
speed advantages.
There are several factors which make it fast. One is that it uses a small constant
(0.1) added to the output derivative when back propagating error. This is useful
because the derivative o(l — o) is 0 at d € {0, 1}. It allows for error to propagate
through these points. It also uses the hyberbolic arctangent function to propagate
the error back. For small error this is linear, but for large error it approaches infinity.
It is never allowed to be larger than plus or minus 17. This contributes to larger
step sizes.
The algorithm uses information about the previous step size and direction. It
also uses two risky assumptions. The first is that the error versus weight curve can
be approximated by a parabola whose arms open upward. The other is that the
change in the slope of the error curve, as seen by each weight is not affected by
all the other weights that are changing at the same time [2]. In calculating new
weights, for each weight the previous and current weight error slopes and the weight
change between the points where the slopes were measured are used to determine a
parabola. Then a jump is made directly to the minimum point of the parabola.
The above is the general idea behind the bigger step sizes in Quickprop. The
technique clearly has some possible difficulties, like taking infinite step sizes. For
complete information on how this problem and others are dealt with the reader is
referred to Fahlman’s paper [2].
It is necessary to note that like standard back-propagation algorithms, there is
no guarantee that Quickprop will converge for any given problem. Convergence,
for example on the exclusive or problem, will often depend on the choice of initial
weights. In Fahlman’s experiments it did not diverge significantly more than back-
propagation.
90-9
V Representation
A key to be able to effectively apply machine learning techniques is the representation
given to the problem [10], Three different representations have been used in this
study with mixed results. The first two methods are geometric hashing schemes .[7],
They are designed to provide some invariance to the fact that the range profile will
not always be aligned in any standard fashion, can be corrupted by noise and will
be compressed and expanded due to changes in aspect angle. At this point in should
be noted that this study has involved planes flying level. That is they have a zero
elevation angle. Only changes in aspect angle are considered here.
The=first representation scheme tried was a geometric hashing scheme. The 256
input points were binned into 64 bins each of 4 points. The bins were normalized
so that the minimum bin served as 0. Next the peak points were found. Hashing is
then done. Master and slave points are chosen. The master will be placed at 0 the
slave at 1 and the other binned points^placed in scale as appropriate. The output is
then normalized so that the maximum height is 1. This provides 127 inputs, 63 to
the left of 0 and 63 to its right.
The second representation scheme was a modification of the first. Now the data
is binned into 51 bins of points each. A cosine curve is placed over each of the peaks.
These curves are added together in order to capture the overlap of peaks. Hashing
is done as before except that the points are not shifted to the left of the master.
This provides us with 51 inputs.
Depending on the number of masters and slaves chosen, each of the above tech¬
niques will produce more than one hashed pattern for a given radar image. In
fact the second scheme will often produce over 100 patterns, where the first might
produce a few more than 20.
The third representation scheme initially normalizes the data to be between 0
and 1. The lowest value returns are set to 0 and the highest set to 1 with the others
appropriately scaled. Then bins of size 10 are made ( currently ignoring the last six
points which are usually zero, but must be used in general). An average is calculated
for these bins and the 25 average values serve as the input to the, learning algorithm.
This scheme is clearly simpler than the first two and does not take into account the
fact that the scatterers must be lined up. However, it is simply an attempt to find a
representation which is a good discriminator. In the results section it will be shown
that this representation has some very good characteristics.
VI Results
In this section, we will discuss the results from both SC-net and Quickprop. The
results will be presented by the representation with which they were derived. This
is also the chronological order in which this study was carried out.
a Results from the first Hashing Representation
The first geometric hashing method was initially used with data on a Lear and
F16 around the 0® aspect angle. The Lear and F16 were chosen since it was felt
they would be clearly distinguishable. It was necessary to determine the number
of hidden layers and the number of cells in those layers. It is known [5] that one
hidden layer is all that is necessary in order to approximate arbitrary functions in
algorithms like Quickprop. Hence, we have used one hidden layer, three total layers,
in all experiments. The results are summarized below in Table 1. Training is done
at 0 degrees.
The training examples were clearly distinguishable. The error reported is the
mean square error over the outputs. In this case, there are 12 outputs. They consist
of an angle from 0 to 45 at 5 degree increments and the two possible planes. From
the results it is clear that it was difficult to recognize the F16. The really bad part
of this is that the F16 was identified strongly as a Lear (often with an output of 1,
the maximum). It usually got the Lear correct. However, it is unclear if everything
90-11
Table 1: Results on testing Lear vs. F16 around 0 degree aspect angle
Plane
Degree
Number recognized
hidden units
Train error
HHQIQ
F16
16
14
28
12
.000486
251
F16
16
13
28
12
.000442
351
F16
16
14
28
12
.000317
451
F16
16
10
28
8
.00428
350
F16
16
6
28
8
.00712
300
Lear
16
25
30
12
.00114
251
Lear
16
'
26
30
8
.00712
300
Table 2: Results on testing Lear vs. F16 around the 15 degree aspect angle
is being seen as a Lear with just 2 planes.
Since it was felt that around 0 degrees might be an unusually troublesome spot,
we tried some experiments with the Lear and F16 at 15 and 16 degrees. The results
are summarized in Table 2. Again the results are weak on the F16, but good on the
Lear. It is also the case that misclassified F16’s are strongly classified as Lears.
Another experiment was tried it which 3 hashed F16 and Lear patterns were
randomly left out of the train set at 15 degrees. After training they were presented
and ail three Lears were recognized, but two of three Fl6’s were misclassified
(strongly). The other was barely recognized. This furthered our belief that the
90-12
hashing scheme might be a difficulty.
Next, noise was added to the 0 inputs. of which there were quite a few. For any
input below .0002, noise (a value from 0 to 0.05) was:randomly added; This slowed
convergence, as it was done each epoch. After 110 epochs with 8 intermediate nodes,
training the Lear and F16 at 0 degrees, we found at 1 degree only 4 of 16 F16’s could
be recognized. Again they were for the most part strongly assigned as Lears.
With SC-net using the original binned and hashed data several experiments were
conducted. Three different methods of representing the hashed data to SC-net were
investigated. The results of each of the methods are listed below:
Plain Version (No variables):
t Lear and F16 jet trained at 0 degrees aspect angle
Obtained (at 1 degree angle):
92.9 % accuracy in Lear using Lear data. 62.5 % accuracy in F16 using F16
data.
• Lear and F16 jet trained a 15 degrees aspect angle.
Obtained (at 16 degree angle):
93.3 % accuracy in Lear using Lear data. 64.3 % accuracy in F16 using F16
data.
• Lear, F14, F15 and F16 jets trained at 15 degrees aspect angle.
Obtained (at 16 degree angle):
100% accuracy in F14 using F14 data.
Obtained (at 17 degree angle):
75 % accuracy in F14 using F14 data.
Obtained (at 18 degree angle);
100 % accuracy in F14 using F14 data.
Using Fuzzy Variables. This method was used to improve the recognition of the
90-13
E16 at 1 and 16 degrees aspect angle. The following partioning of the fuzzy variables
was used:
veryJow:
0.0- 0.2 (0.0, 0.6)
low:
0.2 - 0.4 (0.0, 0.7)
medium:
0.4 - 0.6 (0.1, 0.9)
high:
0.6 - 0.8 (0.3, 1.0)
very-high:
0.8 - 1.0 (0.5, 1.0)
The above ranges are explained as follows by an example. VeryJow is from 0 to
0.2, but it still has some belief up until 0.6 tapering off to 0 [9], This representation
allows for overlapping fuzziness in the recognition space. Then we trained the Lear
and F16 jet at 15 degree angle:
We obtained at 16 degree angle a 95.8% accuracy in F16 using F16 data. We
obtained at 1 degree angle 56% accuracy in F16 using F16 data.
The third method was to use discrete data to generate rules. By looking at
discrete points of the data it is possible to generate rules. These rules will 100%
correctly classify all the examples in the training set, by using only a small subset
of all the data points. We will next identify some of the advantages in using this
scheme.
• Give symbolic meaning to the radar returns, which may help identify possible
reasons for the problems encountered with the F16 and the Lear jet.
• If the discrete data version gives results similar to the other two versions overall
results and the results are good, rules can be easily ported onto other systems
for practical use (both in speed and hardware requirements when designing an
actual system).
• This method allows compression of the network to a more suitable size, since
many connections and cells will be redundant and can be eliminated, again
90-14
yielding memory and speed improveinehts.
Performance was still weak, making the rul^ of no use. At this point, it appeared
that some change in the representation was needed.
b The second hashing scheme
In this case with the Quickprop system, we tried the F16 and Lear at 15 degree
for the training set. In this case there were 113 Lear patterns and 26 F16 patterns.
The network was unable to reduce the error to a reasonable level. It had most of
its outputs in the train data incorrect well into the training. We then tried out the
same problem with SC-net. It correctly classified the Lears, but mis-classified 5 of
the F16s in the train data as Lears and could not get 15 of the others distinguished
from the Lear. Based on these results and some plots of the hashed data it was clear
that the patterns were being obscured by the hash process.
c The revised representation
It has been the case that the results from Quickprop and SC-net have been very
similar with SC-net providing slightly better results. Most of the results from this
representation are from SC-net. However, we first report on the results from Quick¬
prop.
First we trained with the F16 and Lear from 10 to 20 degrees on the odd degrees.
Testing was then done with the even degree patterns. This resulted in a 100% correct
recognition all 10 of the test patterns. The lowest recognition was 0.93.
The next experiment was with the F14, F15, F16, F18 and Lear. Training was
done at 0, 5 and 10 degrees. Then all the planes at the other degrees made up the
test set. Four hidden units were used and 101 epochs were trained for an error of
.00054. Again recognition was 100% on the 40 test patterns.
The following are the extensive SC-net simulation results from this representa¬
tion. Tests were again conducted using the Lear and the F16 data. Prior to this
90-15
they had presented lots of problems in obtaining correct classifications. The follow¬
ing experiments were conducted:
(a) The Lear and the F16 data was trained on all even points within the range
of 10 and 20 degrees inclusive. Then, both planes were tested on all odd degrees.
Result: 100% accuracy on the unseen points was reached for both planes.
(b) In the second experiment we only trained for both planes at 10, 14, and 18
degrees. We then tested on the remaining degrees.
Result: Again 100% accuracy was obtained. In some cases the belief provided
was lower, but the two planes were clearly distinguishable.
(c) In the next experiment we only trained at 10 and 20 degrees for both planes
and tested on the remaining.
Result: Again 100% accuracy was achieved. In some occasions the certainty
would again be lower, but still the planes were distinguishable.
The above results were extremely encouraging and prompted us to further inves¬
tigate the behavior at the 0 - 10 degree range. In prior attempts it was impossible
to distinguish an F16 from a Lear at 1 degree after it was trained at 0 degree. We
conducted the following experiments:
(d) We trained the network for both planes at 0, 5, 10, 15, and 20 degrees. Then
we tested on all remaining points.
Result: Like in the previous cases we were able to obtain 100% accuracy.
(e) In our final experiment we added the data of an F14. In this test we trained
a Lear, an F16, and an Fl4 at 15 degrees and tested if the F14 could be correctly
classified at 16, 17, and 18 degrees.
Result: Again we obtained 100% accuracy, in fact the results v/ere extremely in
favor of the F14, even though only one training example was provided for the F14.
In the final experiment a set of 7 planes were chosen for simulation. The Srcrcs
simulator allows generation of radar return data for 8 planes one of these planes
being the 747 Jumbo. We have excluded this plane from our study, since it appears
90-16
to be quite different from all the other planes. The following 7 planes were used for
testing:
a) F4 Jet b) F14 Jet c) F15 Jet d) F16 Jet e) F18 Jet f ) Lear Jet g) T38 Jet
To test how good the data was in allowing generalization, we decided to train
for all 7 planes at 0, 5, 10, 15, 20, 25, 30,, 35, 40, and 45 degrees. Testing was on
all other degrees within the range [0,45]. Since there are a total of 322 instances
of data and 70 were used for training, a test and training size of 79% and 21%
respectively were used. In almost all instances a correct response was obtained if
the best choice was selected. Problem points existed at 16 degrees were the system
could not distinguish between a Lear and an FI 5 Jet given the Lear data for testing.
At 17 degrees the problem was between a Lear and a T38 given the Lear data. The
certainties were low in both instances, indicating uncertainty in making the proper
classification. Two additional points of uncertainty were at 41 degrees. This time
SC-net had difficulties to distinguish between an F15 and an F14 given the F15 data,
and between an F16 and an F14 given the F16 data. Counting these four as miss-
classifications SC-net obtained an overall classification accuracy of 98.5%. This is a
favorable result, which is an improvement over results reported in [Ij.
VII Recommendations
The above results have shown that the wide-band data efficiently binned can be used
for effective training with SC-net. It is also shown that with the new representation
performance is good in Quickprop, also. It is expected that the results in Quickprop
will continue to parallel those of SC-net. The new data is rather small (only 25
features) and there exists only one example of a plane for every degree. This stands
in contrast to the prior method which sometimes generated dozens of hash patterns
for every degree. Even though, we were able to generalize from rather different
aspect angles, another problem still remains. The wide-band data provided by the
90-17
simulation tool shows the radar-return of the plane taken from an exact center spot.
Jn reality the image will be shifted. In this Cc^e it is expected that the wide-band
data will also be shifted. This will represent serious problems for the binning.. One
possible improvement of the current algorithm is to start the binning prpc^s from
the highest peak of the return, by making it the center of the newly formed return.
In this method a sort of ordering of the highest returns is achieved. This method
could improve making the generated data be more invariant to the shifting of the
peaks for different points. Additionally the relative magnitude of all the returns will
be preserved, which could lend itself as another factor when deciding what plane is
represented.
It is clear that the representation scheme is very important in the recognition
process. With the good results in the artificial data, we feel that real data should
be examined. It will have the problems of shifting and noise. Additionally, there
will likely be other issues. Examination of it will help serve to guide the research
in the proper direction. The current results are encouraging. They indicate that
recognition may be done with relatively small networks. This will allow shorter
training times and more patterns to be addressed.
We have begun to look at the ARTI data, which is actual data with the planes
disguised. The results are very preliminary. It has been necessary to align the data.
The results have been to get 17 of 19 planes correct in a 3 degree aspect arc around
180 degrees. One training example is used of each of the two planes in the set. Again
a problem is representation and it clearly needs to be investigated further.
The symbolic features of SC-net have been helpful in explaining what is hap¬
pening with the different representations. The ability to use variables allows one to
try applying different meanings to ranges of input data in a simple fashion. Rule
generation can show how inputs interact with outputs, even in cases where there is
no (clear) symbolic meaning. It is felt that these types of features are important for
connectionist models in this (and other) domains. Further investigation along this
90-18
line for both Quickprop and SC-net can be useful.
It can be a sort of luxury, due to time factors, to Use more than one learning
algorithm. However, we have found this valuable. Each has different characteristics
which can bring to light distinct problems with repr^entation, the data or type of
algorithm. It is especially useful in a case where the characterization of the data is
not clear. Of course, the characterization of many learning algorithms is incomplete,
also. The comparisons can also shed light on the algorithms characteristics.
Learning in this domain can be very time-consuming in terms of CPU-time. It
requires good, stable, fast computing facilities. Investigations into methods which
are efficient in the learning process in terms of time and space are also of import
to this research. Quickprop is faster than back-propagation schemes and SC-net
is in turn faster than either of those. However, there is certainly much room for
improvement and research in this area is well warranted.
Intuitively, it would seem that the aspect angle of the plane may be recovered
at least somewhat. Training at every 5 degrees it is unknown where a plane at 3
degrees, for example, would be put. Investigation along these lines remains to be
done. Due to early basic recognition problems it was ignored.
The results of this study provide us with optimism that connectionist models
can be effectively used to solve the problem of recognizing airplanes from their radar
profiles. However, there is significant rc.search in the areas discussed here (and others
that may arise) to be done.
References
[1] Atkins, R.G., Shin, R.T., Kong, .I.A. (1989), A Neural Net Method for High
Range Resolution Target Classification, Report from Dept. Of EE and CS and
Research Lab. of Electronics, MIT, Cambridge, MA.
90-19
[2] Fahlman, S.E. (1988), Faster-Learning Variations on Back-Propagation: An
Empirical Study, Proceedings of 1988 Connectionist Summer School, pp.38-51.
[3] Gorman, R.P. and Sejnowski, T.J. (1988), Analysis of Hidden Units in a Layered
Network Trained to Classify Sonar Targets, Neural Networks, Vol. 1, pp. 75-89.
[4] Hall, L.O. and Romaniuk, S.G. (1990), A Hybrid, Connectionist, Symbolic
Learning System, AAAI-90, Boston, Ma. August.
[5] Hecht-Nielsen, R. (1987), Kolmogorov’s Mapping Neural Network Existence
Theorem. Proceedings of IEEE First International Conference on Neural Net¬
works, San Diego, CA., June.
[6] Hinton, G. (1987), Connectionist Learning Procedures, Artificial Intelligence.
[7] Lamdan, Yehezkel and Wolfson, Haim .1., Geometric Hashing: A General and
Efficient Model-Based Recognition Scheme, Robotics Research Lab, Courant
Institute of Mathematical Sciences, NYU, New York.
[8] Michalski,R. S., Carbonell, J. G., Mitchell, T. M. 1983. Machine Learning: An
Artificial Intelligence Approach. Palo Alto, Ca., Tioga Publishing.
[9] Romaniuk, S.G. and Hall, L.O. (1990), SC-net: A Hybrid Connectionist, Sym¬
bolic System, Tech. Report Dept, of CSE, USF, Tampa, FI. (In review).
[10] Shavlik, J.W., Mooney, R.J. and Towcll, G.G. (1989), Symbolic and Neural
Learning Algorithms: An Experimental Comparison, Computer Sciences Tech¬
nical Report #857, University of Wi.sconsin-Madison, Madison, Wise.
90-20
1990 USAF-UES SUMMER FACULTY RESEARCH PROGRAM
Sponsored by the
AIR FORCE OFFICE OF SCIENTIFIC RESEARCH
Conducted by the
Universal Energy Systems, Inc.
ZIML REP.Q.PT
model EQB, CHAMCIERIZINS h PIREPTIQNAL SQUELER PftgEP
■QFSlShL HETERQPYRE DETECTION SVgTEM
Prepared by:
Academic Rank:
Department and
University:
Research Location:
USAF Researcher:
Date:
Contract Number:
Mohammad A. Karim
Associate Professor
Department of Electrical Engineering &
The Center for Electro-Optics
The University of Dayton
WRDC/AARI-2
Wright-Patterson AFB; Dayton; Ohio
Lawrence E. Myers
28 September 1990
F49620-88-C-0053
i I'
MODEL FOR CHARACTERIZING A DIRECTIONAL COUPLER BASED
OPTICAL HETERODYNE DETECTION SYSTEM
by
Mohammad A. Karim
Aismcs:
This report summarizes the research performed during the USAF-UES
Summer Faculty Research Program. The work involved developing an
analytical model for characterizing a directional coupler based
optical heterodyne (coherent) detection system. The coherent
detection system in question consists of two fiber optic links
carrying respectively optical signal and local oscillator beams
which are then combined by means of a directional coupler. The
directional coupler based heterodyning scheme is compared with
that based on Y-coupler as well as that based on only beam split¬
ter in terms of their signal-to-noise ratios. The current analyt¬
ical and simulation results along with those expected to be
generated through a follow-up mini-grant study would be able to
dictate the design characteristics of the most optimum direc¬
tional coupler based coherent detection system.
91-2
ACKNOWLEDGEMENTS
I wish to thank the Air Force Systems Command and, the Air Force
Office of Scientific Research for sponsorship of this research. I
would also like to thank the Universal Energy Systems, Inc., for
providing administrative support of this program.
The support and encouragement of many people made my research
experience both rewarding and enjoyable. I would especially like
to thank Mr. Lawrence E. Myers and Dr. Paul F. MacManamon of
WRDC/AARI-2, Weight Patterson AFB, and Ms. Song H. Zheng of the
University of Dayton for the valuable technical assistance.
91-3
I.. INTRODUCTION
Coherent (heterodyne) detection (Karim, 1990) is a very
powerful technique for the sensing of optically narrowband radia¬
tion at the quantum noise limit. In particular, optical real-time
beam steering devices such as those based on liquid crystals on
nonlinear refracting material (Karim, 1988 and 198.9) can be
anticipated to bring in target signatures and thereafter the
target signatures can be combined with certain laser transmitter
output at a receiver. Such processing of signals at a receiver is
accomplished by means of heterodyne schemes. The mixing of the
two beams, however, can be carried out in two ways: (a) in free-
space by means of a beam splitter; and (b) using fibers and
either a directional coupler or a V-coupler.
The WRDC/AARI-2 Division of the Wright-Patterson AFB is
currently interested in developing a reliable heterodyne scheme
for mixing steered beam and laser transmitter beam preferably
using fiber based systems. As a first attempt, one may develop
model for characterizing the performance of a directional coupler
based heterodyne system where one may limit the analysis to only
step-index fiber and single mode operation. Furthermore, in this
study, the target signature is now assumed to be an ideal one for
the sake of simplicity. It has already been shown that the field
distribution of optical beams as well as detector uniformity play
significant roles in determining the efficiency of the corre¬
sponding heterodyne system (Cohen, 1975; Fink, 1975; Fink and
Vodopia, 1976) . The problems associated with the multimode Y-
coupler are definitely more cumbersome and, therefore, the corre¬
sponding modeling may involve more difficult: analysis.
91-4
One of my current research interests at the University of
Dayton has been directed towards exploring the various possibili¬
ties of designing and characterizing optical systems for generat¬
ing, transforming, shifting, or modifying laser beams. In addi¬
tion, I had also developed analytical as well as software tools
for verifying liquid crystal based beam agility devices for
WRDC/AARI-2 during my 1988 SFRP work as well as for nonlinear
refracting material based E-0 scanner for WRDC/AARI-3 during my
1989 SFRP work. These experiences have served as motivation for
the current SFRP work that also consists of developing analytical
tools but for characterizing a heterodyne system.
II. OBJECTIVES OF THE RESEARCH EFFORT
There has been an ever growing interest in achieving proper
heterodyning of steered optical beams. Several vendors have
already delivered prototype of several liquid crystal based beam
steering devices to my sponsoring laboratory. One of the ways
that such and other beam steering devices could be used would be
to collect target signature and then mix it with laser transmit¬
ter beam in an optical heterodyne configuration. One of the means
of heterodyne scheme may involve the use of a directional coupler
for efficient mixing of the beams at the detector. My assignment
as the SFRP Fellow was to perform basic research into the theory
and characterizations of a directional coupler based heterodyne
system shown in Fig. 1. The system consists of a combination of a
the beam steering device, an acousto-optic modulator, beam split¬
ters and a variable fiber coupler. The advantage of having such a
set-up is that under suitable operating condition, the SNR may be
91-5
reasonably sniall.
The objectives of this research are three-fold:, (a) to
develop model for describing the performance of the aforemen¬
tioned coherent detection system; (b) to develop theory to iden¬
tify which of the two possible couplers (Y-coupler versus direc¬
tional coupler) is more desirable; and (c) to develop software
tool for studying the performance of the said system at least for
a single mode.
III. MODEL DEVELOPMENT
In principle, heterodyning at light frequencies is the same
as that at much lower frequencies. For the conventional receiver,
voltages at two different frequencies are applied to a nonlinear
circuit element, such as a diode, which acts as the translator.
Current at the difference frequency is recovered from the output
of the device and is amplified and processed. Similarly, at
optical wavelengths two electric field at different frequencies
are applied to a photodetector, and current at the difference
frequency is recovered from the output. This technique provides a
signal that is an exact replica of the original but moved down to
some lower frequency range in which amplification, filtering, and
detection are easier to achieve.
Fig. 1 shows the system wherein an input light of frequency
fj^ is mixed with a local oscillator beam of frequency f^^. The
background light that may be present is represented by many small
components, spread over certain band, and each having some ampli¬
tude A]3 and frequency fj^. The local oscillator and input signal
can be represented respectively as E^^ = cos WQt and Ej^ = Aj^
91-6
cos where Wj^ represents the angular frequency of beam x. The
total background light can be represented as Ej, = 21 cos Wjjt.
b
Ajj and Wj3 have values for every component of light present in the
background. The current at the output of the square-law detector
is thus given by
i = c (Aj, cos w^t + Ai cos Wj^^ + cos Wjjt)^. (1)
where c is a constant. As a result of the squaring operation, Eq.
(1) will result into cross products between the various terms,
and we shall obtain components involving the sums and differences
of these frequencies. The sum frequencies are much too high to be
passed by the detector and can, therefore, be neglected. The
resulting expression for the current is given by
i = c [(Ai^ + Ai^ + 2. Ab^)/2 + A^Ai cos(Wi - Wi)t] (2)
b
In Eq. (2), all terms resulting from cross products, or beats,
between the^ Aj^ cos Wj^t components and other components have
been neglected. These beats are negligibly small except in unusu¬
al circumstances given that the heterodyne circuit can provide
excellent frequency filtering as well as critical spatial dis¬
crimination. Eq. (2) consists of direct current (dc) terras re¬
sulting from the local oscillator, the signal and the total
background light and is equivalent to i^j^, = c(Pj^ ^1 ^b^ where
is the local oscillator power, Pj^ is input signal power, and
Pjj is the total background power. The cross term represents the
beat between the local oscillator and signal and is the desired
intermediate-frequency (IF) carrier.
91-7
In addition to the do and the desired IF current, there is
the shot noise that accompanies the dc which i& related to P^,
Pj, and Pjj by
(isn)2 = 2 ce(Pi + P^ + P^) (BW)
= (2e2(BW)/hf) (Pi + Pi + Pfa) (3)
where e is the electronic charge, h is Planck's constant, f is
the light frequency and BW is the IF bandwidth.
The thermal noise at the input to the IF. amplifier is also
important in the same way that such noise is important at the
input to the signal amplifier when direct detection is employed.
This noise power (Karim, 1990) is given by FkT(BW), where F is
the noise figure of the amplifier, k is Boltzmann's constant, and
T is the absolute temperature. Obviously, the best heterodyne
receiver performance will be obtained only when the noise figure
F is the lowest.
The IF signal-to-noise ratio SNR may be obtained to give;
SNR » 2 c2piP3^/( {2ce(Pi + Pi + Pjj) + Fkt}BW] (4)
The noise includes total shot noise given by Eq. (3), and thermal
noise. Furthermore, for practical receivers Pi can usually be
made much greater than P^ + Pj^ and, in fact, can be large that
the thermal noise may be insignificant. Under these conditions,
the only noise of consequence is the shot noise produced by the
local oscillator. Thus,
SNR = n*Pi/(hf (BW) ) (5)
where n* is an efficiency factor. The IF signal power is propor-
91-8
tional to the product of local-oscillator light power and signal
light power. For the direct detector, however, the power of the
recovered signal is proportional to the square of the signal
light power only. Thus, when compared to the direct detection
output, the heterodyne schema provides a gain proportional to
V^i*
From these considerations, we see that the heterodyne re¬
ceiver provides several advantages. First, the conversion process
provides gain so that the signal output of the detector may
overwhelm both thermal and detector noise. Second, the heterodyne
receiver provides excellent discrimination against background
noises. And, finally, it allows for the possibility of recovering
both phase-modulated or frequency-modulated signals.
Heterodyne conversion takes place over an area significantly
larger than the order of a light wavelength. The maximum IF
signal is obtained only when the signal and local-oscillator
beams have the same phase relationship over the complete area of
coincidence. Accordingly, the optical phase must be uniform over
the complete wavefront of each beam. This requirement is met only
under the following conditions;
1. The two beams must have the same mode structure (higher-
order modes are undesirable, since they reduce the fiber
bandwidth and introduce signal distortions) ;
2. The two beams must be coincident and, to maximize the
SNR, their diameters nust be equal;
3. The beams must propagate in the same direction;
4. The wavefronts nust have the same curvature; and
91-9
5. The beams must be identically polarized.
The heterodyne receiver may be considered disadvantageous in
that, for optimum performance, it must meet all of the above
requirements we have set forth, some of which are very severe.
The SNR performance of the heterodyne system is influenced
most particularly by n* which may have contributions from several
fiber optics consideration. Accordingly, n* can be expressed as
i® the fraction of the signal power that is
actually coupled to the fiber, n2 is the transmission of the
waveguide, n3 is the coupling efficiency into the photocurrent
component oscillating at IF frequency, n4 is the amount of output
power actually coupled into the detector, n5 is the bend loss
factor and ng is the quantum efficiency of the photodetection
process. However, n^^ is calculated from the product of another
six launching efficiency components n^^^^ ”12 ' ”l3' ”l4» ”l5»
n]_g. The three spot-size related launching efficiencies (Marcuse
1970 and 1977; Kogelnik 1964) are; (a) n^^ influenced by the
variation of Gaussian spot size w^, (b) n22 influenced by funda¬
mental fiber mode spot size Wg, and (c) n2_3 influenced by the
mismatch of two spot sizes while the three geometrical launching
efficiencies (Karim, 1990) are (a) n24 caused by the transverse
offset, (b) n^^g caused by the angular offset, and (c) n^^g caused
by the longitudinal offset. It must be noted that ng is the same
for all heterodyne systems irrespective of whether or not a
directional coupler is used. Thus for proper comparison between
different heterodyne systems, only the evaluation of n*/ng is
enough.
91-10
IVi SIMULATION
A computer program was written to evaluate n*/ng for the direc¬
tional coupler by having first determined ^12' ”l3' ”l4'
ni5/ riig/ ^2, TI2, n^, and n^ (Neumann, 1988). It may be reasona¬
ble to assume that 1x2 to be approximately l since transmission
attenuation can be as small as 0.2 dB/km (Karim, 1990), and like¬
wise n^ can be about 0.965 for an ideal fiber end-face, and ng is
unity since the distance under consideration is very small. Fig.
2(a) shows the plot of n^j^ versus w^/a where a is the core radius
for different values of normalized frequency V, Fig. 2(b) shows
the values of maximum r\-^2 optimum Wg/a versus V and Fig. 2(c)
shows the plot of n3_3 versus w^/Wg. On the other hand. Fig. 3
(a) -(c) respectively shows the three geometry related launching
efficiencies and n-j^g. Fig. 4(a) shows the geometry of a
directional coupler while Fig. 4(b) shows the values of ng versus
wavelength for different values of coupler separation (Digonnet
and Shaw, 1983) . Fig. 5 shows ng for two of the most important
modes LFg-^ and as functions of wavelength (Karim, 1990) .
Finally, Fig. 6 shows the plots of n*/ng versus the three ratios
z^/Zj-, theta/theta^j, and s/Wg for different values of w^/a where
z^, theta, and s are the three offsets, Zj. is Rayleigh distance
and theta^j is the divergence angle.
V. DISCUSSION AND RECOMMENDATIONS
Preliminary calculations made during this work is indicative
of the fact that it is possible to use directional coupler in an
heterodyne system as it may provide high SNRs provided proper
geometrical and beam characteristics have been maintained.
91-11
Before the design, characteristics of the directional coupler
based heterodyne system is finally specified, a few more aspects
heed to be considered and investigated in the fpllow-up research.
The follow-up rtni-grant proposal is expected to address the
following additional concerns:
(i) consider the fact that the detectors are often nonuniform
and it may contribute to detrimental consequences (Fink and
Vodopia, 1976) that needs to be avoided;
(ii) account for the fact that the steered beams may have field
profile that may result in a reduction of SNR (Fink, 1976; Paran-
to, 1988) ;
(iii) account for the fact that wavelength variation will play a
role in determining system efficiencies (Digonnet and Shaw,
1983) ;
(iv) to account for the consequences of using an Y-coupler
(Karim, 1990; Salzman, Sivan, Kapon, and Katzir, 1983) in the
place of a directional coupler;
(v) to account for the fact that a directional coupler may not be
strictly step-index type; and
(vi) to account for the realistic transmission characteristics of
the available directional coupler.
An analysis of the sort described should be substantiated by
experimental verification of the resulting heterodyne system.
VI. REFERENCES
Cohen, S. C. , Heterodyne detection; phase front alignment, beam
spot, and detector uniformity, Appl . Pot . . Vol. 14, pp. 1953-
91-12
1959, 1975.
Digonnet, M. , and Shaw, H. J., Wavelength multiplexing in single-
mode fiber couplers, Appl. Opt., Vol. 22, pp. 484-4'91, 1983.
Fink, D., Coherent detection signal-to-noise ratio”, Appl . Qpt. .
Vol. 14, pp. 689-690, 1975.
Fink, D. , and Vodopia, S. N., Coherent detection SNR of an array
of detectors, AppI. Opt. . Vol. 15, pp. 453-454, 1976.
Karim, M. A. , Low voltage broadband beam steering devices using
liquid crystals. Final Report to UES, AFSOR Contract No. F49620-
87-R-0004, July 1988.
Karim, M. A. , Analytical model of a unique electro-optic beam
scanner. Final Report to UES, AFSOR Contract No. F49620-88-C-
0053, September 1989.
Karim, M. A. , Electro-Optical Devices and Systems . Boston, Massa¬
chusetts, PWS-KENT Pub. Co., January 1990.
Neumann, E. ~G., Single-Mode Fibers Fundamentals . Springer-
Verlag, Berlin, 1988.
Paranto, J. N., Investigation of heterodyne mixing efficiency
variations due to optical misalignments and aberrations . MS
Thesis, University of New Mexico, December 1988.
Salzman, J., Sivan, U., Kapon, E., and Katzir, A., Heterodyne
detection using multimode waveguide Y-couplers, Appl . Opt. . Vol.
22 , pp. 393 1-3934 , 1983 .
91-13
BEAM
91-14
Fig, 1. The optical heterodyne system.
600 nm
000 nin
SICNAI. wavELEf^GTH
700 nm
Fig. 4. (a) the directional coupler; and (b) n3 versus wavelength
for different coupler separations.
Fig. 5. ng versus wavelength for two inodes LPq^^ and
different values of bend radius R.
91-18
ciency Factor
different values of the
1990 USAF-UES SUMMER FACULTY RESEARCH PROGRAM
GRADUATE STUDENT RESEARCH PROGRAM
Sponsored by the
AIR FORCE OFHCE OF SaENTIFIC RESEARCH
Conducted by the
Universal Energy Systems, Inc.
FINAL REPORT
Context Dynamics in Neural Sequential Learning
Prepared by:
Academic Rank:
Department and
University:
Research Location:
USAF Researcher:
Date:
Contract No:
Kevin G. Kirby. Ph.D.
Assistant Professor
Computer Science and Engineering Department
Wright State University
AAAT-1
Wright Patterson AFB
Dayton, OH 45433
Louis Tamburino, Ph.D.
September 30, 1990
F49620-88-C-0053
Context Dynamics in Neural Sequential Learning
by
Kevin G. Kirby
ABSTRACT
A new neural architecture was developed for efficient learning of spatiotemporal dynamics.
This architecture reduces the learning problem to two subproblems: (1) the formation of a "con¬
text" containing compressed input histories, and (2) the classification of context by an associa-
tional algorithm. The first subproblem was handled by introducing a nonlinear dynamical system
into the neural network, which can be a low-connectivity random net or a continuous reaction-
diffusion system. This enables the solution of the second subproblem to become simpler, requir¬
ing only a variant of the classical perceptron learning algorithm. A theoretical framework was
developed in which the learning capabilities were analyzed in terms of finite automata theory. A
computer simulation system was developed and used to show efficient learning of the sequential
parity problem. Further simulations clarified the role of the context subsystem and demonstrated
promising non-connectionist architectures for this problem.
92-2
Acknowledgements
I wish to thank the Air Force Systems Command and the Air Force Office of Scientific Research
‘ for their sponsorship of this research. Acknowledgement is also due to Universal Energy Sys¬
tems for handling the administrative contact with Wright Patterson Air Force Base.
Above all, thanks go to Dr. Louis Tambutino for his work in helping obtain this summer faculty
research fellowship, and for his probing and critical eye as we tried to crystallize our new archi¬
tecture out of some originally vague insights. I also wish to thank Dale Nelson, head of the
AAAT-1 group, for keeping me infonned about the latest research in neurocomputing and chaotic
dynamical systems, which influenced this woik significantly.
92-3
I. INTRODUCTION
One goal of machine learning research is to develop a system which can make observations
of a complex, unknown environment, and form its own internal model to make predictions of the
enviroment’s responses to various disturbances. This is an extension of what is known as the
"system identification problem": if we watch an unknown system long enough, observing inputs
and outputs, how can we reconstruct its internal states? This problem, in turn, is an extension of
the learning problem that has been tackled by neural network researchers since the field’s begin¬
nings: how do we learn associations in space and time?
The Advanced Systems Research Group of the Avionics Laboratory at Wright Patterson Air
Force Base has been particularly concerned with neural network learning for pattern recognition
and control. As one example, the application to damaged systems is crucial. Can an automated
system participating in the piloting of a severely damaged aircraft rapidly learn a model of the
craft’s dynamics to keep it functional? Neural network architectures truly give us special-purpose
systems with autonomy and real-time capabilities. Exciting developments in the field have
extended the technologies for spatial learning to temporal learning. However straightforward
extensions of existing algorithms (such as backpropagation) to the temporal domain incur a large
cost in terms of processing time and hardware resources.
Because of the cost problem, in my research I decided to approach the temporal learning
problem in terms of a problem reduction: how can we fool a neural network into thinking it is
solving a simple spatial problem when in fact it is handling a complex spatiotemporal one? Ironi¬
cally, this line of research leads us into quite unorthodox neural network dynamics.
My research interests concern the role of dynamical systems in inforaiation processing. My
earlier woik on the biophysical simulations supporting a continuous neuron model, and its appli¬
cation to for robot navigation and path optimization, laid the foundation for my summer assign¬
ment. Aside from its benefits in developing a new technology, my work at the Avionics Labora¬
tory helps create a foundation for a clearer understanding of the role of spatiotemporal dynamics
in intelligence.
92-4
n. OBJECTIVES OF THE RESEARCH EFFORT
Currently, the problem of learning associations between temporal sequences of spatial pat¬
terns is one of the most challenging in the field of neural networks. Inspired by the facility with
which real brains accomplish this learning, generalizations of successful connectionist gradient
descent algorithms from the spatial to the spatiotemporal domain have attracted much interest in
the past two years. Yet the advantages of traditional connectionist solutions have become less
clean these solutions are both time consuming and require large amounts of memory and inter¬
neuronal connections. This suggests it is fruitfiil to search for new ideas.
My summer fellowship allowed me to develop a key concept, called context, and elaborate
it into a novel architecture employing both connectionist and non-connectionist elements. This
architecture has successfuUy and efficiently solved simple sequential learning problems, and
work is currently underway on more complex tasks. The idea of context allows us to decompose
the problem of learning spatiotemporal sequences into two subproblems, solvable by a conver¬
gence of new and old technologies.
The overall problem may be stated in general terms as follows. We are given a set of
multi-dimensional input signals and we observe the corresponding output signals [y(t)).
Unlike a simple spatial mapping, the output at time t depends not merely on the input at that time,
but on an undetermined history of earlier inputs. We wish to have our learning machine form a
model that explains this x-»y association. Our idea is first to generate a compressed representa¬
tion of input histories, called context. We do this by sending the input signals to a dynamical sys¬
tem called a context reverberation subsystem, which serves as a repository for information
relevant to the association. This system may be either a sparsely connected network of conven¬
tional linear-threshold neurons, or an unorthodox complex neuron with continuous internal states.
Hardware requirements for these context reverberation subsystems are less stringent than for ordi¬
nary connectionist learning architectures. The remainder of the problem is to use a conventional
learning algorithm to group the current inputs with the context, and associate these with output
patterns. As a consequence, the introduction of the context concept has allowed us to fall back on
well-studied learning algorithms such as the perceptron convergence method.
In the three technical sections that follow, I map out the ideas and summarize the results of
our implementations. The accomplisliments may be summarized as follows.
Accomplishments of the Research Effort 7/2/90 • 9/7/90
(1) The context reverberation architecture was shown to work efficiently on the sequential par¬
ity problem, which is a temporal generalization of the benchmark spatial parity problem.
An interactive simulation environment written in C++ was developed for these experiments,
(2) The connectionist version of this architecture was shown to require only local connectivity
between neuronal units. This pennits a dramatic, order of magnitude reduction in connec¬
tions. Our solution overcomes the high connectivity problem, the chief obstacle in hardware
implementations of traditional neural networks.
(3) We have formulated necessary conditions for the context dynamics to produce efftcient
learning. The essential feature of the context reverberation subsystem is its ability to
discriminate temporally different but spatially similar inputs. This capability stems from the
highly aperiodic nature of the dynamics. We have found a striking correlation between
period length and the efficacy of the subsystem. This allowed us to bring in results on the
theory of phase transitions in random nets to suggest new context reverberation designs.
(4) We have investigated context reverberation subsystems that are modified by simple genetic
algorithms. This allowed us to improve the amenability of the entire system to traditional
learning algorithms, and results more rapid learning times.
(5) We have mathemaUcally elucidated the nature of the context reverberation approach to
sequential learning. When phrased in the language of finite automata theory, the learning
system is modeling the observed system in a sense different from the textbook "simulation-
as-automaton-homomorphism" definition. Instead, it is employing a mathematically dual
notion we call "quotient simulation". These ideas have implications that transcend the
neural domain that motivated them, and deepen the significance of this woric.
(6) We have identified promising non-connectionist technologies for context reverberation sys¬
tems. The discoveries of the sufficiency of local connectivity (item 2, above) and the char¬
acterization of the requisite dynamics (item 3) suggest that a dynamically richer neuron, the
so-called reaction-diffusion neuron, can be employed in our architecture. A single such neu¬
ron would play the role of an entire local connectionist net of several dozen units.
The final section makes recommendations on the continuation of the research programme.
92-6
m. STRATEGIES FOR LEARNING AN UNKNOWN DYNAMIGS
In this and the following two sections, we describe the details of our research programme.
Since many of the ideas are new and unfamiliar, space limitations require us to be somewhat
terse. Subsequent publications will give these ideas their more complete exposition.
The canonical problem in neural network research is the following: Given a finite subset of
the graph of a function f, guess what f is. The subset of the graph we are given is called a train¬
ing set Elements of the domain are spatial patterns. The algorithm must process this set to come
up with a representation for /. The representation can be used to compute /on elements of the
domain not appearing in the training set and this is called "generalization". Often we are
interested in adding a temporal dimension. This could mean real time, or merely the inforaiation
included in sequences of input patterns provided to the system. In other words, an output pattern
is no longer detennined only by an input pattern, but potentially by an unbounded sequence of
previous input patterns. If we let Y and F be our input and output sets, respectively, we pass from
the interpolation of a function /: AT F to the simulation of a dynamical system
5'.QxX Q where Q is the set of states of our observed system. The job of the learning sys¬
tem is to constnict an internal model of this observed system. Training sets are now sequences of
inputs paired with sequences of output. A good simulation of 5 permits good generalization.
Connectionist systems can be harnessed for such computation by using recurrent networks.
Indeed, backpropagation formally generalizes fairly easily to the recurrent case, although such an
extension seems to be fairly demanding of computational resources [19], This is a very active
area of connectionist research [3,16]. Of course, this is by no means the only sense of "temporal"
learning. One can also phrase the problem as one of learning a sequence of actions in a network
with scalar feedback; this is the work of Klopf [13].
Our architecture is sketched schematically in Figure 1. We assume a discrete time scale.
Spatial input patterns arrive on input lines, and are sent to an output layer of conventional linear
threshold neurons, and to an internal subsystem. This internal subsystem has explicit or implicit
recurrent connections, and is used to store the state of the system. Unlike the state units in the
woric of Jordan [5], the state representation is arbitrary. We call this a context reverberation (CR)
subsystem. In the work of Gallant [6], enhancing an earlier model of Rosenblatt [17], a totally
connected net of linear threshold neurons was used for a similar function. The output of a sequen¬
tial system depends on current input plus state. Our architecture captures this dependence in a
very straightforward way; output units receive signals from the input units and the subsystem.
92-7
TIME DELAY CONTEXT REVERBERATION
NEURAL NETWORKS NETWORKS
' i- con»xt<>»
•Met htttoiy
no dock
CLOCK
NEURAL NETWORKS
♦
nohistoiy
•xactdock
Figure 2. The spectrum of context usage.
state space of observed system
Figure 3. The over-representation of states by context.
92-8
The import^t research problem here is this: how can the CR-subsystem send, an effective
representation of input history to the output layer? Let us define context zs the dynamical state of
the GR subsystem. This concept can be clarified by considering a specthun of approaches to con-
nectionist architectures for sequential problems, shown in Figure 2. At one extreme, time-delay
neural networks send exact delayed copies of the input signals to the ouqiut layer. A conventional
learning algorithm learns to produce conect output fiom inputs x(r), x(r-l), • • • x{t-TMAX),
where die input temporal window [t-TMAK..A] is determined in advance. Hence context in this
case exactly corresponds to history. On the other extreme, some sequential problems depend on
time and not on history. A clock neural network calculates outputs finm the input signal \U) cou¬
pled with an encoding of t.
In many problems, we need an unbounded temporal window, plus some access to encoded
time, but do not want to maintain the overhead of a very large number of time-delayed inputs.
The usual approach would be to designate some neurons as state units, and give them recurrent
connections. These connecdons can be learned by "recurrent back-prop", for example. But our
architecture differs at a deep level from that approach. In such recurrent adaptadon algorithms,
the idea is to create a homomorphism fiom the dynamics of the observed system to the dynamics
of the neural net. If S is the dynamics of the system that we want to model (a finite automaton
that computes parity, for example), we would to set up the weights of our neural net so that its
dynamics is given by 5^ where there is a mapping h from external states Q to net states Q' that
preserves the dynamics. This means that these two dynamics are coordinated by the reladon
5'(/i (^),x) = h(5(q,xy) for all inputs x and states q of the observed system. This homomorphism
of finite automata corresponds to the simulation reladon: the learning system is supposed to simu¬
late the observed system.
Our model constructs a representation of the observed system in a different, innovadve
sense. We have a mapping $ that collapses many states of our CR-subsystem onto each state q of
the observed system. This is depicted in Rgure 3. The equivalence classes of states, <()“'(?).
correspond to the ellipses in the bottom half of the figure. The set of these classes is called the
"quotient space" under the mapping ([». It is this quotient space that, as learning proceeds, should
come to represent the known system. As time proceeds, each class flows through the CR state
space Coottom) tracking the transidons in the observed state space (above). In reality, since the
complexity of the CR dynamics is so great, occasionally we wiU find that generalization off a
training set of sequences is poor over long time periods. Referring to the figure, this happens
when the dynamics of the CR system does not track the quotient structure. The trajectories
92-9
wander out of their proper equivalence classes, so the learned internal model would break dowil a
few time steps past the end of the trmning data. A good CR architecmre will make this a rare
occurrence.
IV. LOW-CONNECTTVITY ARCHITECTURES FOR CONTEXT REVERBERATION
Having discussed the role of the CR subsystem, we now turn to its implementation. We first
investigated a connectionist architecture, in which the CR subsystem contained a number of
linear threshold units arranged in a grid, each connected to others within a limited neighborhood.
(This contrasts with the total (i.e., 0{n^)) connectivity in the nets of Hopfield, Anderson, and
others.) Synaptic weights are fixed and randomiaed. We had known from the work of Gallant
and King [6] that a totally connected layer of random hidden units could learn sequential prob¬
lems with some success. But dozens of units in a such a highly connected system will result in
hundreds of wires. Our first step, undertaken with a graduate snident at Wright State University,
showed that for the so-called "robot plan task", low connectivity was more efficient [12). During
the summer, an "autopsy" of the networks (from large sequences of randomly generated trials)
showed that the most successful ones were those that had highly irregular dynamical trajectories.
We can see this if we plot the intensity of the firing states of the CR net versus time, for an arbi-
uaiy clamped input. (Intensity here means we add up the +1/-1 (firing/silent) values.) This is
shown in the two plots at the bottom of Figure 4. The first plot shows a "good" (TR net, which
learned the task quickly, and the bottom a "bad" net that failed to leam. This shows that aperiodi-
city (or, strictly speaking, periods of length much longer than the time scales of interest) is an
attribute correlated with good CTR performance. One way to control this is to adjust the "gain"
signal, which amplifies the input signals coming in to the hidden net. This is shown in the plot at
the top of Figure 4. Low gain yields better results.
The robot plan task, however, merely involves memorizing a set of sequence pairs, and does
not address the issue of generalization. To go beyond this, we had our network leam the sequen¬
tial parity automaton (Figure 6.) The output at time t is the parity of the string of binary inputs
from time 0 to time t. We used connectionist CR nets with various neighborhood sizes, and plot¬
ted the learning times in Figure 5. The perception algorithm was used to make the input/context
associations. The y axis plots the number of perceptron epochs (passes through the training set),
when trying to leam parity from five sequences of duration equal to 15 time units. We used 64
92-10
Figure 4. Periodicity and dependence of period on
Figure 5. Learning times versus connectivity for the sequential parity problem.
0111100001001...
Figure 6. The state transition graph for the sequential parity problem.
92-12
GR units, and averaged over 20 different configurations. We only plot the result for cases when
every instance learning the training data perfectly (zero error). Generalization on a test set of five
random length 15 sequences resulted in an error under 6% for the 7*neighbor case. Preliminary
results from a genetic modification scheme, which remove unchanging context units and swap
weight values, showed modest improvements in learning rates with no significwt change in gen¬
eralization ability.
Figure 5 shows that not only is lower connectivity dramatically mote efficient in terms of
required connections, but it is even more efficient in absolute learning time. The curve’s
minimum occurs at a neighborhood size of 7, corresponding to 7x64 = 448 weights. This can be
compared 64^ = 4096 weights for the Gallant implementation, which took slightly longer to learn
on average. As the neighborhood size shrinks to 6, 5 and 4, learning times increase, but for some
applications weights may be traded off against learning time,
V. NON»CONNECTIONIST ENHANCEMENTS TO THE ARCHITECTURE
We can conclude from the experiments just described that the locality of the connections in
the CR architecture is a feature to be exploited. This is an economic issue: fewer connections
require less space and ease hardware implementations. But it also allows us to more rigorously
investigate the dynamical properties of CR systems. In this section we discuss our work in rela¬
tion to work in the dynamical systems disciplines. Insights from these disciplines are important,
because they help us understand and extend the capabilities of the CR architecture.
Our connectionist CR-subsystem uses a ratjdom network. Kauffman [7] showed that net¬
works of totally connected random boolean units exhibit an exponential growth in limit cycle
length as the number of units increases. (A limit cycle is one period of the state trajectory.) In
other words, a the state trajectories of even a small net are highly aperiodic. This is termed a
"chaotic" phase, as opposed to the so-called "ordered phase" when cycle lengths increase polyno-
mially with the number of units. Reducing from global to local connectivity slows this growth.
With only 2-neighbor connectivity the periodicity grows as VJv , too slowly for effective use as a
context reverberation net. Kiinen [8] studies threshold units with different local connectivities
and shows that whereas low-connectivity systems in which neighbors are chosen randomly have
exponential growth in cycle lengths, neatest-neighbor systems may show linear growth. The
low-connectivity nets studied were the 3-neighbor "honeycomb" lattice, and the self+4-neighbor
92-13
square lattice. The zero-threshold honeycomb lattice ^ows linear growth in cycle.length, and the
zero-threshold square lattice with self-feedback shows exponential growth. Adding, a unit thres¬
hold moves this system back into an ordered phase. Our research suggests that we should seek a
chaotic phase in our context-reverberation systems. So if we are committed to using a local net
of conventional, neurons. the connectivity level (i.e., neighbortiood size ) should be at least 4. For
connectivity less than this, learning should be impossible. We have experimentally conflrmed
this; the curve in Figure S goes to infinity on the left, when connectivity approaches 3.
Given the efficiency of low-connectivity threshold lattice automata as chaotic reverberation
subsystems, we can investigate the potential of continuous local dynamical systems for setting up
context. The general case is a reaction-di^sion equation, of the form
« V • Z)(^)V«(^,r) + R [u(ii,r)] (1)
Here uOt,r) is an excitation signal diffusing across a space with coordinates d. D is the di^sion
coefficient, which may vary across the space, /? is the reaction term, a function of the excitation
level. In the two terms we have the two ingredients necessary for elective context reverberation:
local communication (via the diffusion term), and local computation (via the reaction term), In
one dimension, in analogy to the discrete set of threshold units we used for learning the parity
problem in the previous section), we can compartmentalize the system to create a ring of com¬
partments. The diffusion term with discretized compartments becomes 2 <>[/*(“/““*)• Thissys-
;«*±i
tern was introduced by Alan Turing [18] to study the destabilizing eHect of diffusion in morpho¬
genesis. (Turing used two diffusing signals.) Othmer and Scriven [15] examined how the
dynamical properties of this reaction-diffusion system depended on the topology, studying rings
and lattices in what was a continuous analog to the studies of Kiirten [8] on lattice automata cited
above.
Can we expect such continuous dynamics to improve our CR-subsystems? We claim that it
should be possible by using a neuronal model based on the Turing morphogenesis equations.
This model is called the reaction-diffusion neuron [9,10], a continuous extension of a discrete
linear u.at used in a Darwinian brain model. These neurons take input signals and map them into
continuous gradients, which evolve by reaction-diffusion equations. Gradients are read by spa-
’ally fixed sensors, whose response induces the firing of the neuron. With suitable dynamics one
reaci^on-diffusion neuron can play the role of an entire CR-subnetwork. It is interesting to note
that one of the features that make such neurons amenable to learning algorithms is their
92-14
gradualism property; slight changes in the internal sensor distiibutioh produce slight changes in
the dynamics. Indeed, this is like the case of totally connected random boolean nets, which have
low structural and Lyapunov stability [7]. On the other hand, we may, trade off gradualism for
long cycle lengths, which, as we have seen in the previous section, are strongly correlated with
good learning performance in CR systems. This suggests that a highly a^riodic reaction-
diffusion neuron would be advantageous. We have the extreme case of this in the chaotic
Lorenz-Turing neuron [II], in which trajectories proceed along a strange attractor. Figure 7
shows a plot a a two-compartment LT-neuron, showing the diffusion-caused interference of the
familiar Lorenz butterfly attractor.
Once we allow continuous dynamics to enhance the CR system, we can consider adaptation
of this dynamics. Recall that in the experiments discussed so far, the CR-subsystcm was random
and fixed, except possibly for sporadic localized genetic modification. This allows the context
dynamics to evolve concunently with the representation. We can view this idea of context as
analogous to some phenomena observed in other research. In the continuous case we can have
formation of topological features such as thdse that arise in Turing-type morphogenesis systems,
e.g., the stripes of Meinhardt and Gierer [14]. A well-known neural analog is in the work of
Amari [1] on pattern formation in neural fields. A neural field changes the representation of neu¬
rons from a set of finite units to a manifold of mathematical points. Firing signals do not pro¬
pagate along connections, but spread out along the manifold, governed by equations of the fonn;
X = -M(n,r) + Jw(p,p')/lw(p',0]dii' + input tenns. (2)
The weighting "matrix" >v(p,p,') internal to the neural field is fixed in advance; only the input
weights change (according to a Hebbian algorithm). Efferents from the neural field in a sense use
the acquired patterns as state information, since the topographic arrangement on the field is also a
kind of repository for input history. Let us call the kind of input context in these morphogenesis
and neural field models topographic context. This contrasts with the concept of scrambled con¬
text used by our CR systems to represent history.
Scrambled and topographic context promise to be important notions in the theory of
sequential learning. We have shown that the requirements for a good context-reverberation sub¬
system do not include the high connectivity required for connectionist solutions to other prob¬
lems. This may encourage molecular electronic hardware implementations [4]. Low-connectivity
nets of linear threshold functions with nearest-neighbor topology provide long cycle lengths and
are an effective means for providing scrambled context information to a single-layer learning
92-15
L7-oeaioo (/*«2)
p»21(X) ^Z667 ^10.00 (cmx)
$60^301
Figure 7. Trajectory of a 2-coinpartinent Lorenz-Turing neuron
92-16
algorithm for learning. the p^ty dynamics. We believe that such a result is encouraging for the
study of ndn-symbphc .non-connectionist continuous systems for solving "real" arUficid intelli¬
gence problems. Far from merely providing a new technology for machine learning, the Gl sys¬
tem has produced fertile ideas enabling a more profound understanding of the learning problem
itself.
VI. RECOMMENDATIONS
In conclusion, we have constructed a new neural architecture, called the context reverbera¬
tion architecture, that (1) learns the sequential dynamics of art observed system: (2) learns rapidly
and with accurate generalization; (3) does so with dramatically low connectivity requirements;
(4) uses aperiodic dynamics to represent context; (5) can be sped up by simple genetic algo¬
rithms; (6) suggests a new paradigm for the sequential learning problem; and (7) may be
enhanced with a continuous time non-cotmectionist neural model.
The continuadon of this project requires a woricing simuladon of a CR system solving a
highly complex sequendal problem. It also demands a M theoredcal characterizadon of the
computadonal power of the architecture. Toward this end, the following recommendadons are
made.
(1) The learning system should attempt to learn a finite state system with a dynamics two ord¬
ers of magnitude more complicated than the parity problem. This means moving from two
states in the observed system, to over a hundred states. With convendonal architectures,
this would be intractable. Experimentadon with this more complicated system will require
an enhancement of the simuladon environment. Figure 8 shows a screen from the cunent
environment.
(2) The connecdonist version of the context reverberadon subsystem should be coupled with a
version of the Doya-Yoshizawa algorithm [2] to get the most computadonal work out of the
context units. This algorithm is a condnuous time extension of recurrent backpropagation.
A detailed study of the tradeoff between this algorithm’s cost and the efficiency gain is in
order. This would smooth the transition to handling continuous input and output signals.
(3) The coupling of the Lorenz-Turing neuron to the learning system should be perfected.
Mathematical analysis done during the summer suggests the Doya-Yoshizawa algorithm
can be modified to act between compartments in a reaction-diffusion system like the
Lorenz-Turing neuron. This ties in to fundamental research on the computational potential
of chaos.
An SFRP follow-on grant from the AFOSR Research Initiation Program will be sought to pursue
these recommendations. We expect contributic^^’ln both technological and theoretical areas.
A M K I
iHwsflO* n>qaw4 w<wnu ourwMO» 9ow>w w«iywt
TASKS w-90fily
H^NOLOCY; mnactiomit
In^ IMMK 1
CMtM( unilK S4
tMgiieofheoa mzit *
Owl^t uniU: 1
TroMnq aat mz« S
Random woiqM onaO; 38378
lnaut-(o~«an(«it 9am t.OO
Stqnoi duraderK t2
Vta trqWhq «oo<n«: SO
Iniomal MooUdon: non«
adoptatlon: p«ckt(td qaretatron
EPOCH
17
0.9
0.8
0.7
0.8
0.3
0.4
OJ
0.2
0.t
0.0
Figure i. A .screen from the AMICI interactive simulation program for CR systems
92-18
REFERENCES
1. Amari, S., "Dynamics of Pattern Formation in Lateral-Inhibition Type Neural Fields," Biological Cyber¬
netics. 1977, Vol. 27, pp. 77-87.
2. Doya, K. and S. Yoshizawa, "Memorizing Oscillatory Patterns in The Analog Neuron Network," Proc.
IEEE/inns Conference on Neural Networks, January 1990, pp. I27-I32.
3. Giles, Ci., G.Z. Sun, H.H. Chen, Y.C. Lee, and D. Chen, "Higher-Order Recurrent Networks and
Grammatical Inference," Neural Information Ih'ocessing Systems 2, D, Touretzky, Ed., San Mateo,
Cab'fomia, Morgan-Kaufmann, 1990, pp, 380-387.
4. Hong, F.T. "The Bacteriorhodopsin Model Membrane System as a Prototype Molecular Computing Ele¬
ment," Biosystems. 1986, Vol. 19, pp. 223-236.
5. Jordan, M,I„ "Serial Order A Parallel, Distributed Processing Approach," Institute for Cognitive Sci¬
ence Report 8604, University of California, San Diego, May 1986.
6. Gallant, S.I., and DJ. King, "Experiments with Sequential Associative Memories," Cognitive Science
Society Conference, Montreal, August 1988.
7. Kauffman, S,A., "Principles of Adaptation in Complex Systems," In Lectures in the Sciences of Com¬
plexity, D. Stein, Ed., Addison-Wesley, 1989.
8. Kiirten, K.E., "Dynamical Properties of Threshold Automata with Nearest-Neighbor Interactions on a
Regular Lattice," Proc. IEEE International Conference on Neural Networks, August 1988, pp.
I37-I43.
9. Kirby, K.G., and M. Conrad, "The Enzymatic Neuron as a Reaction-Diffusion Network of Cyclic
Nucleotides," Bulletin of Mathematical Biology. 1984, Vol. 46, pp. 765-783.
10. Kirby, K.G., M. Conrad, and R. Kampfner, "Evolutionary Learning in Reaction-Diffusion Neurons,"
Applied Mathematics and Computation, To appear, 1991.
11. Kirby, K.G., "Information Processing in the Lorenz-Turing Neuron," Proc. IEEE Engineering in
Medicine and Biology Conference, Molecular Electronics Track, November 1989, pp. 1358-1359.
12. Kirby, K.G., and N. Day. "The Neurodynamics of Context-Reverberation Learning," Proc. IEEE
Engineering in Medicine and Biology Conference, Moleculjir Electronics Track, November 1990,
to appear.
13. Klopf, A.H., "A Neuronal Model of Classical Conditioning," Psychobiology. 1988, Vol. 16, pp. 85-
125.
92-19
14. Meinhardt, H., and A. Gierer, "Generadon and Regeneiadon of Sequences of Structui^ During Mor¬
phogenesis,” Journal of theoretical Biology. 1980, Vol. 85^ pp. 429-4S0.
15. Othmer, H.G., and L£. Scriven, "Instability and Dynamic Pattern in Cellular Networks,” Journal of
Theoretical Biology. 1971, Vol. 32, pp. 507-537.
16. Pineda, FJ., "Recunent Baclquopagadon and the Dynamical Approach to Adapdve Neural Computa-
don,” Neural Computation. 1989, Vol. 1, pp. 161-172.
17. Rosenblatt, F., Principles of Neurodynamics, Washington, DC, Spartan Pre^, 1961.
18. Turing, A.M., "The Chemical Basis of Morphogenesis," Phil. Trans, Royal Soc. B 1965, Vol. 237, pp.
37-72.
19. Williams, RJ„ and D. Zipser, "A Learning Algorithm for Fully Running Recurrent Neural Networks,"
Neural Computation. 1989, Vol. 1, pp. 270-280.
92-20
1990 USAF-UES SUMMER FACULTY RESEARCH PROGRAM/
GRADUATE STUDENT RESEARCH PROGRAM
Sponsored by the
AIR FORCE OFFICE OF SCIENTIFIC RESEARCH
Conducted by the
Universal Energy Systems, Inc
Prepared by: Richard E. Miers
Academic Rank: Associate Professor
Department and Physics
University: Indiana University/Purdue University at
Fort Wayne
Research Location: WRDC/AARI-2
Wright Patterson AFB
Dayton, OH 45433
USAF Researcher: Paid F. McManamon
10 Aug 90
F49620-88-C.0063
Date:
Contract No:
Fiber Laser Preamplifier for Laser Radar Detectors
by
Richard E. Miers
ABSTRACT
A study was made of the feasability of using a fiber laser preampUfier as a
means of improving the detectability of laser radar signals. Although fiber
laser amplifiers at the wavelength of interest, 1.064 pm, have not been
developed, a study of the development of Er-doped fiber laser amplifiers for
1.55 pm indicates the usefulness of such amplifiers. Also the properties of
Nd-doped fibers indicates that such fibera should amplify 1.06 pm wavelength
signals as well as or better than the Er-doped amplifiers. Recommendations
for development and testing of such an amplifier are given.
93-2
Ackaowledgeingnte
I wish to thank the Air Force Systems Command and the Air Force Office of
Scientific Research for sponsorship of this research. The help in all
administrative and directional aspects of this program by Universal Energy
Systems is acknowledged.
The fidendhness and helpfulness of the personnel of AABI-2 are greatly
appreciated. Paul F. McManamon provided enthusiastic support for my
efforts. His advise and help were invaluable. I would like to thank Captain
Larry Myers for providing the initial idea for this project and for his
assistance in gathering valuable information. Carol Heben and Lt. Deanna
Won provided information and help that was quite useful for development of
this project.
93-3
I. INTRODUCTION
The Electro-optics Division of the Avionics Laboratory at Wright- Patterson
Air Force Base is involved in the development of laser radar systems. One
possible method of increasing the detectability of a rettiming laser radar
signal might be to use a fiber optical laser preamplifier immediately before
the photo detector.
In the competition between signal and noise, immediate amplification of the
signal eliminates detector noise sources as areas of concern (due to the higher
signal level after amplification). Fiber laser amplifiers may allow immediate
signal preamplification.
My research interests in recent years have involved the use of various lasers
in atomic physics research. As part of this research I have constructed lasers
and have used a number of different types of commercial lasers. My
knowledge and experience involving lasers contributed to my assignment to
the Electrooptics Division of the Avionics Laboratory.
93-4
IL OBJECTIVES OF THE RESEAECH EFFORT:
Light amplification by stimulated emission was first accomplished with the
advent of Maiman’s Ruby Laser in 1960. [1] Light amplification of a
traveling light wave in an active fiber was reported by Koester and Snitzer in
1964. [2] They meastired the amplification of a 1.06 pm wavelength signal in
a one meter long Nd-doped fiber that was pumped with a flash lamp situated
at the center of the spirally wound fiber. In 1969 Holst and Snitzer detected
1.06 pm signals with a fiber laser preamplifier. [3] They suggested use in a
laser radar system.
Subsequent to the initial efforts noted above, very little was reported on
traveling wav<* fiber laser amplifiers until around 1985. However, doped
fibers have been used in fiber lasers. In 1973 Stone and Burrus observed
lasing in two types of Nd-doped silica fibers. [4] One fiber consisted of an
active core of fused silica, aluminum oxide, and Nd-oxide surroimded by a
fused'silica cladding. The second was made up of an active core of fused
silica and Nd oxide only. In both cases the Nd content was less than 1% by
weight. They were end-pumped with a chopped argon-ion laser beam and
with a pulsed dye laser operating at 590 nm. The first fiber oscillated at 1.06
pm and the second at 1.08 pm. In a later paper they reported cw operation of
Nd-doped fiber lasers at room temperature. [5] In 1979 they reported a self-
contained LED-pumped single-crystal Nd:VAG fiber laser. [6] Since 1985,
due to the interest of the communications industry, a great deal of
development and testing has been done on Er-doped fibers to amplify light
signals at wavelengths at or near 1.55 pm. However, as of now there seems
93-5
to have been no effort to develop traveling wave fiber laser amplifiers at 1.06
^m which is the wavelength of immediate interest for laser radar at the
Avionics Laboratory.
My assignment as a participant in the 1990 Summer Faculty Research
Program (SFRP) was to determine the potential for increased laser radar
receiver sensitivity based upon immediate optical signal preamplification
through use of a Nd-doped fiber laser amplifier. Results of this investigation
show that this type of laser preamplifier may be useful. Therefore the deveh
opment of an experimental program to build and study such an amplifier is a
secondary goal. It will be proposed that this investigation be continued with
funding from the Research Initiation Program.
III.
Fiber laser amplifiers seem to have a number of advantages over other optical
amplifiers. Fibers have been developed to transmit a single mode signal.
These fibers are polarization independant. Laser diodes with wavelength 807
nm, which is within a very efficient absorption pump band for Nd^+, are
readily available for end pumping . Theoretically, it appears possible to
develop a Nd<doped fiber amplifier with ideal i.e. quantum limited signal-to
noise ratio.
/
93-6
Kingston derives the signal-to-noise ratio for a laser amplifier operating in
single mode applied to both coherent and direct detection systems. [7] His
treatment is similar to that of Yariv. [8] They show that the signal*to*noise
ratio for an ideal 4>level laser amplifier can be given by
(S/N)p*Py(hVtf.> (t]
where (S/N)p is the signal-to-noise ratio for a single mode of amplified
spontaneous emission, P, is the power of the detected signal, hilis the photon
energy at the center of the stimulated emission band,iiyis the detector
bandwidth, h is Planck’s constant. This is the ideal quantum limited receiver
equivalent to a heteradyne detector with quanttun efficiency of 1.
Nd-doped fibers are four-level laser systeros while Er-doped fibers are three-
level systems. In a four-level laser amplifier, population inversion should
exist at very low absorbed pump powers since the lower lasing level is
initially unpopulated. In the three-level amplifier the lower lasing level is at
or near the ground state, therefore the ptunping rate must be high enough to
move at least half of the electrons from the lower level to the upper level
before amplification can occtir.
Many of the techniques and experiments leading to the development of Er-
doped fiber laser amplifiers may be used in the development of Nd- doped
amplifiers at 1.06 pm. Also a futiure goal at the Avionics Laboratory may be
to develop laser radar systems using the eyesafe wavelength of 1.55 pm. In
that event an Er-doped laser preamplifier may be used.
Yamada et al. have obtained a signal gain as high as 37.8 dB in an Er-doped
fiber amplifier pumped by 0.98 pm laser diodes. [9] They used a 30 m long
93-7
Er®'*’ doped fiber at 20 mW launched pump power. The pumping source was
an InGaAs strained quantum well LD that could deliver up to about 85 mW
output. They foimd by comparison that 0.98 pm pumping was more efficient
than 1.48 pm pumping. Giles et al. measured the noise performance of an
Er-doped fiber amplifier pumped at 1.49 pm. [10] They achieved a gain of
37.4 dB at a wavelength of 1.53 pm and a noise figure of 5.3 dB. The
sensitivity of the optical receiver without preamplifier was -21 dBm. The
receiver sensitivity with preamplifier was -41 dBm. Olshansky defines the
noise figure F as 2X where^is the population inversion factor, given by
y ■ sMt- fls)
where g, is the rate of stimulated emission and a, is the rate of stimulated
absorption. [11] For complete population inversion, ya 1, and the noise figure
FisddB.
End pumping has been accomplished using various pumping lasers and
wavelengths. In particular pumping with diode lasers has been achieved. To
date laser diode pumping has been reported at 0.807 pm, 1.48 pm, and at 0.98
pm with the piunping at 0.98 pm reported to be the most efficient. Yamada,
et al. tested the noise characteristics of Er^*** doped fiber amplifiers pumped
by 0.98 pm and 1.48 pm laser diodes. [12] The spontaneous-spontaneous beat
noise and the spontaneous shot noise for the 1.48 pm pumping were higher
than those for the 0.98 pm pumping. The noise figures estimated from the
signal-spontaneous beat noise were 3.2 dB for 0.98 pm pumping and 4.1 dB
for 1.48 pm pumping. Bour, et al. describe a 980 nm diode laser suitable for
pumping Er^'*' fiber amplifiers. [13] These lasers are capable of supplying 125
mW of CW power. Takada et al. measiured picosecond laser diode pulse
93-8
amplification up to 12 W using laser diode pumping at 1.48 jjm. [14] In
another test of an optical receiver with an Errdoped fiber preamplifier, an 8.3
dB improvement of the signal-to*no^.over the r^eiver without the
preamplifier was recorded. [15] The Er>doped fiber had a 7.2 pm core, 17 m
length, and 300 ppm Er^'*' concentration in pure silica fiber. The signal was
detected with an APD/FET receiver.
An important issue related to optical amplifiers is the phenomenon of
amplification of spontaneous emission (ASE). This occurs concurrently with
signal amplification and degrades the signal-to*noise ratio (SNR). Some of
the features that affect this are dopant concentration distribution across the
fiber, the pump and signal mode spatial overlap, the signal absorption due to
the lower level population, the amplification of spontaneous emission, the
detailed structure of the fluorescence and the absorption spectra, and the
gain homogeneous and inhomogeneous broadening. Excited state absorption
which is known to occur at some pump wavelengths may also have an effect.
These features have been investigated both theoretically and experimentally
for Er-doped single-mode fibers. [16] For long fibers it has been shown that
end-pumping codirectionally with the signal provides lower ASE at the
detector than does pumping contradirectionally. [11][17]
Optical amplification characteristics as a function of length and doping
concentrations are important. These effects have been investigated for Er-
doped silica single-mode fibers. [18] It was fotmd that the increase of the Er
concentration causes deterioration of the amplification characteristics in the
Er-doped silica glass system, even when the Er concentration was less than
93-9
loop ppm. This relationship is very important in the optimum design of the
Er-doped fiber amplifiers.
A number of detectors are available at 1.06 pm. Silicon detectors have the
advantage of low noise at room temperature, however 1.06 pm is at the edge
of their useful detection range which lowers the quantum efficiency.
Germanium PIN avalanche photodiode;s are probably the most sensitive
detectors at this wavelength, however the noise at room temperature may be
considerably higher, InGaAs detectors lie somewhere between these two in
both sensitivity and room temperature noise. The quantum efficiency of
silicon detectors at 1.06 pm is quite low, around 10% or less. The quantum
efficiency of Ge or InGaAs detectors can be quite high, however the dark
current at room temperature is much higher. Since a fourdevel laser
amplifier such as Nd^'*' doped fiber can be theoretically equal to the ideal
quantum limited detector with quantum efficiency of 1, tise of such an
amplifier could make a direct detector as sensitive as the ideal heteradyne
detector if the filtered bandwidth can be kept as low as that for the hetera¬
dyne detector.
Many of the papers of Desurvire and others examine the theory of Er- doped
fiber amplifiers much of which seems to be verified by experiment. [19][20]
IV. RECOMMENDAnONS
Noise generated by a laser fiber amplifier is generally due to quantum or shot
noise generated by amplified stimulated emission. Researchers appear to
have come close to bringing it near the quantum limit for Er-doped fiber laser
93-10
amplifiers. It seems reasonable to assume that similar techniques applied to
the development of Nd-doped fiber amplifiers would produce efficient noise-
iree gain at the 1.06 pm wavelength.
Digonnet and Gaeta have theoretically analyzed an optically pumped fiber
laser amplifier. [21] They show that for an ideal 4-level amplifier the
unsaturated single pass gain factor can be given as
where the gain is given by
(V
lout/lin is the ratio of amplified signal to the input signal, a is the stimulated
enussion cross section for the amplified wave, tf is the flourescent lifetime of
the upper lasing level, is the photon energy of the p\unp light, is
the effective intensity of the absorbed pump light m the fiber.
Applying this theory to a Nd-doped silica fiber assuming a* 3.x 10'^ cm^, tf ■
4.5 X 10’* s, hl^(800 nm) ■ 2.48 x 10’^® J, » 3 x 10“* cm) « 3 x lO’’’^
cm^ gives a slope efficiency of
V/P.b, = 0.18/mW or G<dB)/P.b, = 0.78 dB/mW.
Af is the cross-sectional area of the pumped fiber of radius r.
Po, et al. reported a slope efficiency of 0.47 dB/mW in a Nd-doped silica fiber
when pumped at 800 nm. [22]
93-11
Since doped fibers will probably have to be purchased from some manu¬
facturer, the absorption and emission spectra may already have b^n
investigated and published as technical specifica:tions. If sp^al fibers have
to be manufactured for test purposes, it may be n^essary to make
measurements of these properties to evaluate the wavelengths at which they
should be pumped and the probable gain profile of the particular fibers. One
coinpany, 'York 'VSOP, (Chandler’s Ford, Hants, UK) is manufacturing single¬
mode fibers in which 300 ppm of Nd ions have been incorporated. The Gain
bandwidth centers at 1088 nm. These fibers probably would not be useful as
amplifiers at 1060 nm. Fibers have been made that have bandwidth centered
at 1060 nm. [4][22][23][24][25] These are co-doped with AI2O3. Researchers
at Rutgers University and at Brown University have the capability of
manufacturing custom doped fibers that have amplification centered at 1060
nm and have indicated the willingness to provide fibers for test purposes. A
number of manufacturers are making erbium doped fibers and may have the
facilities to custom dope other fibers. These options are being investigated at
this time.
The gain of a fiber amplifier will be a fimction of pump wavelength and
power, fiber length, doping, efficiency of pump coupling into the fiber, and
spa cial overlap of pump and signal modes. In a four-level laser medium the
population inversion and therefore the gain factor is proportional to the pump
power absorbed. On the other hand, for a three-level system such as Er-
doped glass, the gain factor is only positive when the pump intensity exceeds
the threshold intensity for which the upper level population exceeds that of
the lower or ground level. Therefore for any given doping concentration and
93-12
power level there m an optimum length for maximum gain :n an Er*doped
fiber. The same is not true for the four-level Nd-doped system. In such a
system the fiber can be arbitrarily long in order to absorb as much pump
power as possible. Nominal tmpumped lengths should have an insignificant
eff^ on signal absorption. These are properties that will need to be studied’
in the development of a Nd-doped fiber amplifier.
To ascertain the usefulness of a fiber amplifi,er, the most important aspect
will be to study the sensitivity of the amplified detector compared with the
sensitivity of an unainplified detector Parameters ithat have to be studied are:
gain, signal, reflections, methods of launching pump light into the fiber,
methods of coupUng the input of the signal to be amplified into the fiber,
methods of coupling the amplified output into the detector, and methods of
reducing the detected bandwidth. All of these factors could affect the signal-
to-noise ratio of the amplifier. These are all aspects that investigators of Er-
doped fiber amplifiers have studied.
To build a Nd-doped fiber laser amplifier a single mode fiber doped to amplify
signals at 1.06 pm wavelength will be needed. A suitable diode laser at 800
nm for pumping the fiber is needed. Also needed are either a pigtail coupler
or a dichroic mirror with reflecting and transmitting properties such that
both the pump light and signal light can be coupled simultaneously into the
fiber. Positioners for precise coupling and coupling eyepieces would be
needed. A photodetector for detecting 1.06 pm light will be needed along with
a narrow band filter centered at 1.06 pm to reduce the spontaneous emission
noise and to eliminate any unabsorbed pump light.
93-13
Equipment needed for studying the properties of the amplifier are a power
meter for measuring input and output signal and pump light, A signal laser
at 1.06 pm, and attenuators for vaxying the signal.
One method of ascertaining the center of si^al gain band of the fiber is to
place appropriate mirrors at either end of the fiber and measure the
wavelength at which it ia^s or oscillates. Also this may provide an
appropriate signal laser for testing the gain of the fiber laser amplifier.
Other miscellaneous test equipment will include an appropriate optical table
or breadboard for mounting, a monochrometer for measuring wavelengths of
signal and pump light, and motmting equipment for mirrors, lenses etc.
93-14
REFERENCES
[1] T. H. Ms^an, Nature 187, 493 (1960); Brit. Conunun. and Electr. 7,
674(1960)
[2] C. J. Koeater and E. Snitzer, "Amplification in a Fiber Laser,” Applied
Optics, vol. 3, pp 1182-1186, 1964.
[3] C. G. Holst and E. Snitzer, "Detection 'with a Fiber Laser Preamplifier
at 1.06 p.,” IEEE J. of Quantum Elect., vol. QE-5, pp 319- 320, 1969.
[4] J. Stone and G. A. Burrus, "Neodymium-doped silica lasers in end-
pumped fiber geometry,” Appl. Phys. Lett., vol. 23, ^pp 388-389, 1973.
[5] J. Stone and C. A. Burrus, "Neod3rmium-Doped Fiber Laser: Room
Temperature cw Operation with an Ipjection Laser Pump,” Appl.
Optics, vol. 13, pp 1256-1258, 1974.
[6] J. Stone and C. A. Burrus, “Self-Contained LED-Pumped Single-
Crystal Nd:Y^G Fiber Laser,” Fiber and Integrated Optics, vol. 2, pp
19-46, 1979.
[7] R. H. Kingston, Qptisfll.and Infrared Radiation, “
Springer-Verlag, (Berlin), 1978, Ch 8.
[8] A. Yariv, Optical Electronics. Third Ed., Holt, Rinehart and Winston,
(New York), 1985, Appendix D.
[9] M. Yamada, M. Shimizu, T. Takeshita, M. Okayasu, M. Horiguchi, S.
Uehara, and E. Sugita, “Er^+ -Doped Fiber Amplifier Pumped by 0.98
^m Laser Diodes,” IEEE Photon. Technol. Lett., vol 1, pp 422-424, 1989.
93-15
[10] C. R. Giles, E. Desurvire, J. L. Zyskind,.and J. R. Simpson, ^dise
Performance of Erbium-doped Fiber Amplifier Pumped at 1.49 ^m, and
Application to Signal Preamplification at 1.8 Gbits/s,” IEEE Photon.
Technol. Lett., vol 1, pp 367-369, 1989.
[11] R, Olshansl^, “Noise Figure For Erbiiim-Doped Optical Fibre
Amplifiers,” Electron. Lett., vol. 24, pp 1363-1365, 1988.
[12] M. Yamada, M. Shimizu, M. Okayasu, T. Takeshita, M. Horiguchi, Y.
Tachikawa, and E. Sugita, “Noise Characteristics of Er®+ doped Fiber
Amplifiers Pumped by 0.98 and 1.48 m Laser Diodes,” IEEE Photon.
Technol. Lett., Vol 2, pp 206-207, 1990.
[13] D, P. Bour, N. A. Dinkel, D. B. Gilbert, K B. Fabian, and M. G. Harvey,
“980 nm Diode Laser for Pumping Er®'*’ Doped Fiber Amplifiers,” IEEE
Photon. Technol. Lett., Vol 2, pp 163-155, 1990.
[14] A. Takada, K Iwatsuki, and M. Saruwatari, “Picosecond Laser Diode
Pulse Amplification up to 12 W by Laser Diode Pumped Erbium-Doped
Fiber,” IEEE Photon. Technol. Lett., vol 2, pp 122- 124, 1990.
[16] N. Henmi, Y. Aoki, S. Fujita, Y. Sunohara, and M. Shikada, “Rayleigh
Scattering Influence on Performance of 10 Gb/s Optical Receiver with
Er-Doped Optical Fiber Preamplifier,” IEEE Photon. Technol. Lett., vol
2, pp 277-278, 1990.
[16] E. Desurvire and J. R. Simpson, “Amplification of Spontaneous
Emission in Erbiiun-Doped Single-Mode Fibers,” IEEE J. Lightwave
Tech., vol 7, pp 835-845, 1989.
[17] E. Desurvire, “Spectral Noise Figure of Er^'*’ - Doped Fiber Amplifiers,”
IEEE Photon. Technol. Lett., vol 2, pp 208-210, 1990.
93-16
[18] M. Shimizu, M. Yamada, M. Horiguchi, and E. Sugita,
**Concentation Effect on Optical Amplification Charactmstics of Er<
Doped Silica Single<Mode Fibers,** IEEE Photon. Techhol. Lett., vol 2,
pp 43-45, 1990.
[19] E. Desurvire, *‘Analysis of Erbium-Doped Fiber Amplifiers Pumped in
the ~ ^^13/2 Band,*' IEEE Photon. Technol. Lett., vol 1, pp 293-296,
1989.
[20] E. Desurvire, J. R. Simpson, and P. C. Becker, "High-gain Erbium-
doped Traveling-wave Fiber Amplifier,** Optics Lett., vol 12, pp
888-890, 1987.
[21] M. J. F. Digonnet and C. J. Gaeta, "Theoretical Analysis of Optical
Fiber Laser Amplifiers and Oscillators,** Appl. Optics, vol 24, pp
333-342, 1986.
[22] H. Po, F. Hakimi, R. J. Mansfield, B. C. McCollum, R. P. THimminelU, E.
Snitzer, "Neodymium Fiber Laser at 0.905, 1.06 and 1.4 pm, "Abstracts
of Annual Meeting of Optical Soc. of America, Seattle, WA (1986) Paper
FD4, p 103.
[23] P. R. Morkel, M. C. Farries and S. B. Poole, "Spectral Variation of
Excited State Absorption in Neodymium doped Fibre Lasers,’* Optics
Comm., vol 67, pp 349-362, 1988.
93-17
[24] E. Snitzer, H. Po, F. Hakimi, R. P. IHimminelli, B, C. McCollum,
"Double Clad, Offset Core Nd Fiber Laser,” Optical Fiber Sensor Conf.
Postdeadline Paper, New Orleans, LA (1988) Paper PD5.
[25] H. Po, £. Snitzer, R. P. TVumninelli, L. Zenteno, F. Hakimi, N. M. Cho,
T. Haw, "Double Clad High Brightness Nd Fiber Laser Pumped by
GaAlAs Phased Array,” Optical Fiber Communication Conf.
Postdeadlline Paper, Houston TX (1989) Paper PD2.
93-18
1990 USAF-UES SUMMER FACULTY RESEARCH PROGRAM
GRADUATE STUDENT RESEARCH PROG^
Sponsored by the
AIR FORCE OFFICE OF SCIENTIFIC RESEARCH
(AFOSR)
conducted by the
Universal Energy Systems, Inc.
FINAL REPORT
Reusable Ada Software -
Evaluating the Common Ada Missile Packages (CAMP~3)
Prepared by:
Academic Rank:
Department and
University
Research Location:
USAF Researcher:
Date :
Contract No
Brian J. Shelburne Ph.D.
Associate Professor
Mathematics and Computer Science
Wittenberg University
WRDC/AAAF-3
Wright-Patterson AFB
Dayton, Ohio 45433
Marc J. Pitarys
August 17, 1990
F49620-88-C-0053
Reusable Ada Software -
Evaluating the Common Ada Missile Packages (CANP-3)
by
Brian J. Shelburne
ABSTRACT
One of the largest and earliest projects involving reusable
Ada software was the United States Air Force sponsored CAMP
effort with McDonnell-Douglas Corporation. This summer's AFOSR
project evaluated CAMP for its usefulness and suitability for
avionics applications.
During the process of evaluation, errors were discovered in
some of the CAMP softvare parts. The tight dependencies among
the various CAflP parts caused by "wlthing" and the poor internal
documentation made tracking down these errors extremely
difficult.
CAM? is overly complex, poorly documented, and contains
errors. The final conclusion arrived at is that CAMP software is
not suitable for avionics applications.
94-2
Acknowledgements
I wish to thank the Air Force Systems Command (Avionics
Laboratory) and the AFOSR for sponsorship of this research.
Universal Energy Systems must be mentioned for their concern and
help to me in all administrative and directional aspects of this
program.
I wish to especially recognize Marc Pitarys for his guidance and
assistance with this project and to thank him for the many and
fruitful discussions we had on the problems and pitfalls related
to software reusability. I also want to thank Kenneth Littlejohn
and James Williamson both for their assistance with my project
and for the many fruitful discussions we had.
I. INTRODUCTION;
A concise explanation of CAMP found on page 195 of Appendix I of
Developing And Using ADA Parts in Real-Time Embedded Applications
states
"The main goal of the CAMP program has been to
establish the feasibility and value of reusable Ada
software within the mission critical real-time domains.
This has required a careful evaluation of a particular
domain f the development of reusable components, the
development of automated support for software reuse in
the spftware djvelopmeht lifecycle, and the application
of both reusable components and the automated tools to a
realistic application."
There are three phases to the CAMP project. Again, to quote from
the above docunent
"(Tihe CAMP program began in 1984, with a 12-month
feasibility study. There were two major objectives; (1)
to determine if sufficient commonality existed within
the nissile operational flight software to warrant the
development of reusable software parts; and (2) to
deteimine che aspects of parts engineering that could be
fully or partially automated, and to develop the
requ;rements and top-level design for a parts
conpositior; system to support reuse".
"While CAM?-1 concentrated on feasibility analysis.
Phase 2 of the CAMP program (CAMP-2) was a 32-month
technology demonstration phase that began in September
1985. The goal of CAMP-2 was to demonstrate the
technical feasibility and value of reusable Ada missile
parts and a PCS (Parts Composition System] by building
and using tlem on a realistic application."
It was during the CAMP-2 phase that 454 software parts were
identified and citaloged from 10 different missile systems. A
prototype parts composition system (called AMPEE or Ada Missile
Parts Engineering Expert system) was constructed to manage these
parts for later reuse. The CAMP parts and the AMPEE system were
94-4
then used to construct the navigation and guidance systems of an
11th missile application which was tested in a MIL-STD 1750A
ha rdwa re-in- the-loop simulation. At the same time a suite of
"armonics" benchmarks were developed from the CAMP parts to
measure the effectiveness and efficiency of Ada compilers for
armonics applications.
The third phase of CAMP (CAMP-3) was technology transfer. The
CAMP parts were extended to over 500 in number and the AMPEE
parts composition system was re-engineered, rewritten in Ada, and
renamed the Parts Engineering System (PES). a user's manual.
Parts Engineering System Catalog User's Guide - version 1.1 was
also written to support it.
II. OBJECTIVES:
The objectives of my research were threefold: learn the CAMP
Parts Engineering System, use it to generate a sample avionics
application, and evaluate the usefulness of CAMP.
III. THE PARTS ENGINEERING SYSTEM (PES)
The CAMP Parts Engineering System (PES) is used to access the
CAMP parts database. It has two basic functions : submission of
new parts into the database and user examination of parts
currently in the database. A part is an Ada package, subroutine
or task, often a generic. The exact definition is unclear bat
one criterion for a part is that it must be able to stand alor.e.
94-5
Parts can and frequently do "with" other parts.
PES is menu driven and easy to use. It has some built-in helps
facilities,. The written documentation for PES is well written.
PES accesses the parts database by letting the user generate a
search list of parts to examine. There are 24 different search
criteria (e.g. part ID, part number, part name, keyword, etc.)
with a limited capacity to generate coapound search queries using
Boolean AND's and OR's. Once a list of parts is selected, the
user has the option of examining any of the 22 attributes of a
part in the current search list or examining the source code of
the part (specification or body). Examples of attributes are
abstract, keywords, classification, projects used, by, sample
usage etc. Final selection of a part for use generates a small
text file containing the specification and body file names plus
compilation instructions.
III.l. Critique of PES ;
The limited screen capabilities of the VT220 terminal made PES
slow and awkward to use particularly with the way it generated
new screens. A window-like eivironment would be better.
Of the 24 different search criteria, only two, search by keyword
and search by classification, were useful. Of the 22 attributes
attached to a part, only three, the abstract, compilation
instructions and sample usage, were useful. In many cases the
abstract was too sketchy to be of much use.
94-6
For a given part, one could either examine the specification file
or the body file. Unfortunately CAMP grouped many different
parts into one specification file so a search of the
specification file had to be made to find the part specification.
Fortunately PES used a VAX editor to access CAMP part files so
one could use the "find" command to locate the specification.
Unfortunately, many body files use the Ada "is separate" feature
to shift the part code into a second file making the code
inaccessible from PES. The only way to find the code was to exit
PES, do an operating system level search (such as the directory
command) for a file name that was similar to the body file name,
and then access that file outside of PES. Fortunately CAMP'S
file naming scheme made such a search easy.
Each CAMP part source code file was highly structured in term of
comments and a standard format for comments was followed for all
files. Many of the comments vere not particularly useful (e.g.
the entire revision history) ifhile other comments that would have
been useful were missing (e.c. the mathematics underlying the
algorithm used to implement the part). There was no
documentation on what conditions must exist before a subroutine
was called or on what conditions were true after a subroutine
completes. In any case, the large quantity of comments tended to
get in the way and long searches were often necessary to find
code. There were ilsc errors in t)ie comments, generally minor
annoying but they could cause problems if one was lot familiar
with the parts database or with that part's function. PES also
referenced perts whose files did not exist; sometime; PES simply
94-7
had the wrong file name.
One final and very annoying feature of the specification and body
files was that they were often more than 80 columns wide so they
would not fit on an 80 column screen.
The most telling fact about PES was the way I eventually got
around having to use it. I made a hardcopy of the "taxonomy" of
the parts data base which gave me a classification of the CAMP
parts and a general overview of all parts. (This taxonomy was
used by PES as the classification search criterion.) I then made
a hardcopy listing of the directory of specification files. The
standardized file naming conventions for CAMP parts made this
easy to do. Then as needed for each specification file, I made a
hardcopy listing of the body files for that specification.
Fortunately CAMP file names were meaningful. In this way I could
examine the CAMP parts with an editor without using PES. Since
the part attributes provided by PES were in the source code as
comments, I stiJl had access to them.
It is significant that a simple way around PES was found that
allowed easier and better access to the parts database.
IV. BUILDING APPLICATIONS WITH CAMP PARTS
The main goal was to study the suitability of CAMP parts for
avionics software. To accomplish this goal and to get a feel for
using CAMP pacts, three small programs of increasing complexity
and a final avionics application were written.
94-8
The three small programs were
1. A square root program using CAMP'S math routines
2. A program to calculate actual north and east velocity given
nominal north and east velocity and the wander angle
3. A program to calculate great circle distance between two
points given their latitude and longitude.
The final avionics application was a very simple "waypoint
navigation program. It accepted a sequence of coordinates
(latitude and longitude) for the starting point, waypoints, and
terminating point of the route of an aircraft then computed the
great circle distance for each segment of the route and the
turning angle at each waypoint.
The idea for the last two programs and the avionics application
were suggested by various C/MP parts.
IV. 1. Square Root Program
The program was to read a real number, use the CAMP part to
calculate the square root, and then print i:t. Since the
specification of the CAMP square root part stated that the
exception "Negative_Input' would be raised if the argument to the
square root function was negative, an exception handler was
written.
Unfortunately or negative input, the "Negative__input" exception
was not raised as it should have been. Instead a "Constraint Error
exception was raised which crashed the program.
94-9
This program revealed the first serious problem with CAMP; that
is, the complexity of using the GAMP parts.
CAMP parts are, for the most part, generic packages and
procedures. Many of them "with" other parts and the resulting web
'of interconnections and dependencies can be complex and
confusing. Tracing down the origin of the square root constraint
error required going back through seven files only to discover
that CAMP used the standard VAX Ada square root function; It did
not reveal why the ''Negative_Input" exception was; not raised.
The second of the seven files, General_Purpose_Math, contained
the line "package body Square_Root is separate". The third of
the seven contained the line "separate General__Purpose_Math" thus
bridging the link from the second to the third file.
Unfortunately, there happened to be ah (unknown) eighth file
which also contained the line "separate General_Purpose_Math"
and this was the one tiat was used. There was no indication
that this other file existed and indeed the file was found by
accident only after J was forced to cast around for some other
outside "source" of the error. The file contained a different
method to implement a square root function and its documentation
stated 'The exc(5ption "Negative_Input" ; is raised if "Input" is
negative'. TJis was false; the code failed to do this.
The unknown iighth file was used instead of the third file
because it was apparently compiled later. This brings up a
second problem with CAMP parts aside from the heavy use of
■Vithiig" and the difficulty of back-tracking problems; that is
94-10
the use of Ada's "separate" clauses that lead to configuration
management abjuses like having multiple "separate" files. The
third problem is simply a lack of clear concise part dependency
documentation. Fourth, there is no warning that if one uses a
part, it won't generate a tangled web of dependencies. Finally,
using hidden code in implementing a part increases its complexity
and the likelihood of an error.. The part user has no knowledge
of and no control over this.
IV. 2. NE Velocity Program
The main work for this program was done by two CAMP generic
procedures contained in the package
Wander^Aaimuth__Navigation_Parts. Both generic procedures had
three type parameters and two subprogram parameters.
In designing parts, CAMP use's an approach known as the "semi-
abstract data type" method, a method based on generics and use of
overloaded operators. It means that many CAMP parts are generics
that are tailorable to user defined data types but with CAMP
providing a set of default types and operators. (An excellent
description of the Parts Design Alternatives is found in Appendix
III of the CAMP manual Developing And Using Ada Parts in Real
Time Embedded Applications. )
Instantiating the generic package required the user to declare
three data types and two subroutines as generic parameters. This
insures that strong typing, one of Ada's best features, is
maintained but it requires more work to instantiate the generic.
However, CAMP provides a part called Basic_Data_Types which
;proviaes many of the necessary data types and operators that
operate on them. To assist the user, each CAMP part documents a
"sample usage" on how to use the default data types for that
pant.
The NE Velocity program was written using the sample usage
examples given in the source code since the easiest way to write
the program was to follow CAMP'S suggested guidelines. However,
the program would not compile due to an "inconsistency detected
during overload resolution" error when trying to find an actual
parameter corresponding to one of the generic formal procedure
parameters.
The problem is somewhat complicated to explain but it has to do
with instantiation of generics. According to the Ada Language
Reference Manual, if a generic package is instantiated and a new
type is derived from the instantiation then subroutines from the
instantiation ought to be visible with parameters of the derived
type. This does not seem to be the case.
The problem once understood was easy to fix but either the CAMP
documented sample usage is wrong or there is an Ada compiler
problem. (This example was run on two different compilers and
the same error occurred. The CAMP manuals did mention problems
with immature compilers instantiating generics.) Whether the
error was actually a compiler problem or the result of incorrect
code was never determined.
IV. 3. Great Circle Distance Navigation Program
94-12
This program implemented the CAMP package Great_Circle_Arc_Length
and again the documented sample usage was used to guide the
implementation.. Using the lesson learned from the NE Velocity
program, the "overload resolution inconsistency" problem was
avoided.
However, when run, the program sometimes crashed with a con-
straint_error raised by a "Modified_Newton_Raphson" routine for
square root. Use of a debugger eventually traced down the error
to a calculation of 1.00000000000000000000001295525 for a square
root (obviously using quadruple precision) where the output value
was constrained to the range -1.0 .. 1.0. which generated the
constraint error.
Again because of the tangle of "withed" parts and part
dependencies, the error was very difficult to trace. Because a
lot of parts were hidden, the occurrence of this particular error
came as a complete surprise.
The error did reveal a fundamental design flaw : a square root
routine whose design guaranteed only a certain amount of
precision (which was not documented) was "misused" by permitting
higher precision arithmetic internally.
V. WAYPOINT NAVIGATION APPLICATION
The Waypoint Navigation Problem : Given the earth coordinates
(latitude and longitude) of three points A, B, and C, find the
great circle distances AB, BC, and the turning angle at point B.
This particular application was chosen because CAMP had the parts
that would make the implementation fairly easy. The one drawback
was that the mathematics behind the particular method CAMP used
to compute great circle distances and turning angles was not
documented. The mathematics had to be extracted out of the code
which again was made difficult by the complex web of CAMP part
dependencies. Understanding the math behind the CAMP parts was
important because the exact meanings of the part parameters was
not well documented.
Unfortunately the CAMP package subroutines for Waypoint
Navigation contained a fundamental error in the method used to
calculate turn angles. Thus the application never worked.
VI . EVALUATION
CAMP is overly complex, poorly documented, and contains errors.
The complexity of the CAMP package derives from heavy usage of
"withing" which generates a complex web of part dependencies.
This is complicated by use of Ada's "is separate" which means
that a part might be scattered over three or more files.
The complex web of part dependencies has three drawbacks. First,
tracking down an error where there is no indication of what file
the part is in or even what files the part depends on is
difficult, time consuming, and frustrating. Second, the
ramifications of the use of any part are unknown. The user has
no feel for how or why the part works since too much is invisible
94-14
to him or her. Third, it makes for more complicated and slower
running code if a part is implemented using calls to other parts..
Complicated code is more susceptible to error.
Another aspect of the complexity of CAMP parts deals with the
"semi-abstract data type" method of parts design. Many CAMP
parts were generic units which were tailorable to user-defined
types. Using a part required defining a number of data, constant,
and subroutine parameters to be used in the instantiation of the
generic. While use of CAMP'S default types took some of this
burden off the user, it was not clear why all were needed or even
why certain distinctions are made. Default types always added to
the complex web of dependencies between parts. In all fairness
to the CAMP authors, they took the best way out of a difficult
situation. While generics aid software usa^^ility, Ada's strong
typing requirements are hostile to it.
Complexity of part dependencies and loss of visibility into the
inner workings of a part decreases user confidence in the part.
While one may argue that a specification is all that is needed to
use a part, any responsible programmer would at least want to
check out the code to be assured that it works.
The specification of a part should explain clearly, concisely,
and exactly what the part does. It should state the constraints
required on any parameters and it should guarantee the precision
of any output. Only with this will a user have confidence in it.
The CAMP parts fail to do this.
In general the comment documentation was well laid out and
followed a standard format but there was too much useless,
documentation. The mass of documentation got in the way when one
was trying to find a particular code segment.
Finally there were erro.rs in the CAMP parts. Each of the four
programs written uncovered problems ranging from errors in the
parts documentation to serious mathematical design errors.
VII. CAMP AND REUSABLE SOFTWARE
My experience with CAMP was not one of success. I found too many
errors that in many cases were difficult to track down. I lost
confidence in CAMP.
For software reusability to succeed, a user must be able to use a
reusable part having the assurance that it will take less time to
implement and that it will be correct. The parts and all its
inner workings must be totally visible to the user so that he or
she knows the exact behavior of the part and so that he or she
can be assured that the part works correctly. The part must be a
white box, not a black box, in order that the user may have
confidence in it. CAMP parts, by and large, are black boxes.
Regarding software reusability, there is one fundamental issue in
that must be addressed. When a software engineering project
is undertaken, choices are made at all levels of the design
process from the highest level of abstraction to the lowest level
of implementation. Two designers given the same specification
will come up with two equally good designs yet at some point the
94-16
designs will diverge and from that point become increasingly
incompatible so that it is impossible to reconcile one design
with the other.
This raises the issue of whether it is even possible to take the
software parts from one design domain (armonics) and use them in
a second design domain (avionics).
As I became more and more acquainted with CAMP parts I began to
perceive the underlying design choices that were made. When I
wrote my programs I had to reconcile my own design tendencies
with the design of CAMP parts. In each instance I had to ask the
question^ will my design take less time and be more assured of
being correct or can I use a CAMP part and achieve the same
degree of software quality in> less time?
VIII. RECOMMENDATIONS
My recommendations address five issues ; Parts Engineering
Systems, documentation of software parts, quality of software
parts, Ada reusable software, and software engineering issues.
VIII. 1. PES Issues :
A Parts Engineering System supporting a reusable software library
should use a windows-like interface. Hypertext might be a good
approach .
A selection process should be used that allows the user to
eventually narrow his or her search to the examination of a few
94-17
parts. Selection criteria used to search for parts should be
useful and meaningful.
A PES should display the salient information about a part so that
the user can decide quickly whether to investigate it further or
reject it.
VIII. 2. Documentation
Documentation for a software part should state the exact
constraints required on parameters and it should guarantee the
precision of output. The algorithm or implementation used by the
part should be explicitly stated or a reference given for it.
Documentation of sample usage should be included. Dependencies
on other parts should be explicitly stated. The user should not
be overwhelmed by useless documentation. Quality documentation
is a necessary precondition for a Parts Engineering System.
VIII. 3. Quality
Reusable software, parts should be totally visible to the user so
that the user may understand its function, proper application,
and correctness. Reusable software parts should be white boxes,
not black boxes.
Reusable software parts should not be overly dependent on other
parts. In addition, any dependency should not go back more than
one level. All dependencies should be clearly stated in the
documentation.
94-18
VI 1 1. 4. Ada and Reusable Software Parts
The use of Ada's "is separate" feature should be avoided.
The grouping of many software parts into a single file should be
avoided.
Reusable software is supported by Ada generics but inhibited by
Ada's strong typing. CAMP'S "semi-abstract data type" based on
the generic and overloaded methods (see page 212-213 of
Developing and Using Ada Parts in Real-Time Embedded
Applications ) is a reasonable compromise provided software part
dependencies are minimized and well documented.
VIII. 5. Software Engineering
Software reusability across different application domains might
not be possible. If different designers create diverging designs
from the same specification, divergence will be much greater with
different specifications.
Users of reusable software from a particular applications domain
should be proficient in that domain in order to allow the rapid
evaluation of the usefulness and correctness of a part. Human
judgment should be part of the loop to insure the part performs
the task correctly.
REFERENCES
Biggerstaff, T.J and Perils, A.J. ed, Software Reusability - Vol.
I Concepts and Models; ACM Press, New York, 1989.
Developing And Using Ada Parts in Real-Time Embedded Applications
( CAMP- 3 ) ; McDonnell Douglas Missile System Company, 1990.
Parts Engineering System Catalog User^s Guide Version 1.1;
McDonnell Douglas Missile System Company, 1990.
94-20
1990 USAF-UES SUMMER FACULTY RESEARCH PROGRAM/
GRADUATE STUDENT RESEARCH PROGRAM
Sponsored by the
AIR FORCE OFFICE OF SCIENTIFIC RESEARCH
Conducted by the
Universal Energy Systems, Inc.
FINAL REPORT
COMPUTER SIMULATION OF NMOS INTEGRATED CIRCUIT CHIP
PERFORMANCE INDICATORS
Prepared by:
Academic Rank:
Department and
University ;
Research Location:
USAF Researcher:
Date :
Ashok K. Goel, Ph.D.
Assistant Professor
Department of Electrical Engineering
Michigan Technological University
Electronics Technology Laboratory
WRDC/ELE
Wright-Patterson Air Force Base
Dayton, OH 45433-6543
Dr. Fritz Schuermeyer
July 20, 1990
Contract No:
F49620-88-C-0053
COMPUTER SIMULATION OF NMOS INTEGRATED CIRCUIT CHIP
PERFORMANCE INDICATORS
by
Ashok K. Goel
ABSTRACT
For an integrated circuit chip based on the silicon NMOS
technology, a computer-efficient model of the various chip
performance indicators has been developed and a user-friendly
computer program called ’’NCHIPSIM" suitable for the simulation
of the chip performance indicators for an NMOS microprocessor
or a gate-array chip has been developed. In addition to
predicting the various chip performance indicators such as its
maximum clock frequency, power consumption, computational
capacity, power efficiency, fabrication yield, functional
throughput rate and the size of an NMOS chip with the given
technology parameters, the program NCHIPSIM has also been used
to simulate the dependence of the various chip performance
indicators on the technology feature size in the range 0,1-
2.5 microns and the chip integration level in the range 100-
100,000,000 transistors on the chip. The results have been
compared with and found in excellent agreement with those
known for several single-chip microprocessors based on the
silicon NMOS technology.
95-2
ACKNOWLEDGEMENTS
First, : like to thank the Air Force Systems Command,
the Air Force Office of Scientific Research and the
Electronics Technology Laboratory at the Wright-Pattersoh Air
Force Base for sponsoring this research. I also wish to thank
the Universal Energy Systems, Inc. for helping me with all the
administrative and directional aspects of this program.
I also like to thank several individuals who helped in
making this e.xperience, truly rewarding and enriching for me.
First, I am grateful to Dr. Fritz Schuermeyer for providing
me with support, encouragement and a truly enjoyable working
atmosphere. Second, I am thankful to Mr. Don Peacock for his
personal attention at several occasions during this work. The
concern of Dr. Ben Murphy was greatly appreciated. The help
of Mr. Darrell Barker, Mr. Ben Carroll and Mr. Tim Seiter was
invaluable ir. overcoming several technical roadblocks.
95-3
I. INTRODUGTiGN:
It is- extremely important to be able to simulate, the
performance of an integrated circuit chip before its
fabrication is undertaken. This can be accomplished by
executing the following steps:
a) Development of a computer-efficient model of the various
chip performance indicators;
b) Development of a user-friendly computer program suitable
for the chip performance simulation; and
c) Application of the computer simulator for the determination
of the performance indicators for a chip with known values of
the various technology parameters.
Continuous advances in the integrated circuit technology
have resulted in more complex chips integrating millions of
devices and interconnections. In the recent years, it has
become necessary to use interconnections in two or more levels
to achieve higher packing densities, shorter propagation
delays and smaller chips. Further, because of the much higher
mobility of electrons in Gallium Arsenide (GaAs) , it has
emerged as a preferred substrate for the development of the
high speed circuits.
The Device Technology Branch of the Electronics
Technology Laboratory at the Wright Patterson Air Force Base
is interested in the development of very high speed integrated
circuits based on the GaAs FET and other technologies. In
954
particular, they are concerned about the parasitic effects
that adversely affect the performance of a high-speed high-
density integrated circuit. My previous research work has been
focused on the contribution of .the parasitic effects on the
crosstalk and the propagation delays in the GaAs MESFETs and
the high-density interconnections on the GaAs-based very high
speed integrated circuits and this contributed to my present
summer assignment.
II. OBJECTIVES OF THE RESEARCH EFFORT:
The long range objective of this research is to develop
the computer-efficient algorithms and the related user-
friendly computer software modules suitable for the simulation
of the performance indicators for the small-geometry high¬
speed high-density integrated circuit chips based on the
Silicon NMOS, Silicon CMOS, Silicon BJT and the Gallium
Arsenide FET, HBT and other high-speed technologies.
It was decided that, first, a computer-efficient model
and a user-oriented computer program would be developed to
predict the various performance indicators of a silicon NMOS
chip sich as its maximum clock frequency, power consumption,
computational capacity, power efficiency, fabrication yield,
functional throughput rate and its size. In addition to
simulating the performance of a single-chip microprocessor or
a gate-array with the given technology parameters, the program
95-5
should enable the user to analyze the effects of scaling the
feature size or changing the integration level of the chip on
its performance indicators. In order to validate the algorithm
and the simulator, the simulation results will be compared
with those for several existing single-chip microprocessors
based on the silicon NMOS technology.
III. ALGORITHM DEVELOPMENT:
To date, an NMOS chip performance simulator called
"NCHIPSIM" has been developed. In the algorithm, first, the
average interconnection length on the chip in units of the
average logic gate dimension is determined by the given
number of transistors or gates on the chip and the Rent's
constant by using the equations derived by Donath (1) . Then,
for a logic intensive interconnection-capacity limited VLSI
chip, the average logic gate dimension on the chip is
obtained by setting the interconnection available per gate
equal to that required per gate. Then the value of the average
interconnect length on the chip is evaluated and the values
of the chip size and the chip area were determined. Next, the
average delay time in a gate on the chip is obtained by
adding the various component delays such as the delay due to
the output resistance of the driving gate and the interconnect
capacitance, the delay due to the input capacitance of the
gate at the next stage, the distributed-RC delay of the
95-6
interconnections and the delay due to the resistance of the
interconnections and the input capacitance of the gates. Next,
the total delay suffered by an input signal on the chip is
determined by adding the delay through the logic
gates, the contribution of the global interconnection delay,
and the contribution of the speed-of-light limit depending on
the propagation speed cf the electromagnetic waves on the
chip. The maximum cloc< frequency of the chip is then
determined. Next, the power consumption of the chip is
calculated by adding the power consumption in the logic gates
and the dynamic power consumption at the I/O buffers. This
depends on the power supply voltage, fraction of the on-chip
gates that switch during a clock period, the total capacitance
at an output pin and the number of pins ^per chip as determined
by using the Rent's rule [2] or provided by the designer.
Next, the computational capacity of the chip is determined
by using the number of gates on the chip and its maximum clock
frequency. Then the power efficiency of the chip is obtained
by using the values of che computational capacity and the
power consumption. Next, the functional throughput rate of the
chip is obtained by using the values of the computational
capacity and the chip area. Finally, for a known value of the
density of defects on the chip, z\q fabrication yield of the
chip / is obtained by using the ?r;ce law [3] . The flowchart
of the progran "NCHIPSIM" is shov/n on the following tv;o pages.
95-7
START
Calculate Interconnection Resistance,
Interconnect Capacitance, Gate Resistance,
Number of Gates and Pins on the Chip
Calculate Average interconnect Length, <
Gate Dimension, Chip Size and Chip Areal
Calculate Typical Gate Delay andl
Maximum Clock Frequency
Calculate Power Consumption, Power Efficiency,!
Throughput Rate and Fabrication Yield
Write Results on Screen!
and NCrilPSIM.OUT i
Redefine^
Reference
Chip? ^
YES
(ContiPuec C'' \ex: ^ac9)
95-8
I at j
^Tyes
input Scaling Factor and
Density of Fabrication Defects
Calculate Scaled Transistor
and interconnect Parameters
Calculate and Write Results for Scaled
Chip on Screen and NCHIPSIM.OUT
More Scaled
^hip Simuiatim
New IntegratiorT''--^
Input Number of Transistors
and Density of Defects I
YES
ICalculate and Write Results
Ion Screen and NCHIPSIM.OUT
YES
More IntegratioTr
Levels? _ -
STOP
FLOW CHART FOR 'NCHIPSIM*
95-9
IV,. THE PROGRAM "NGHIPSIM";
Using the flowchart and the steps outlined in the above:
section, an extremely flexible and user-friendly program
called "NCHIPSIM” suitable for the simulation of the various
performance indicators for an NMOS integrated circuit chip
with known technology parameters was written in FORTRAM-77 and
run successfully on the VAX mainframe computer. First, the
program allows the user to choose the chip type, a
microprocessor or a gate array. Depending on the chip type,
the program lists the constants, found empirically, required
for determining the average interconnect length and the number
of pins on the chip. Next, the program lists, but allows the
user to change interactively, the values of several chip
parameters such as the number of transistors, fan-out of a
typical gate, capacitance at the output pin, density of
fabrication defects and the probability of an on-chip gate to
switch during a clock period. Next, the program lists the
typical values of the transistor related parameters such as
its feature size, input capacitance, output gate resistance,
power supply voltage and the ratio of optimum-size to feature-
size transistors but permits the user to change any of these
values for the chip being simulated. Next, the user can choose
a dielectric material out of silicon dioxide, polyimide,
alumina and epoxy glass or select one of his own and define
its dielectric constant interactively. Next, the user can
95-10
choose an interconnection material out of aluminum, copper,
silver, tungsten and molybdenum or select any other material
and define its electrical resistivity interactively. Next, the
program lists the typical values of the other interconnection
parameters [4] for the technology feature size defined earlier
but allows the user to modify any of these values
interactively. These parameters include width, pitch and
thickness of on-chip interconnects, thickness of the
dielectric material, number of interconnection layers and the
utilization coefficient of the interconnections. Then, for the
chip defined above, the program calculates and displays the
values of its performance indicators such as its size, the
maximum clock frequency, power consumption, computational
capacity, power efficiency, functional throughput rate and the
fabrication yield. Next, the program allows the user to scale
the reference chip defined above by a certain scaling factor
and determines the performance indicators for the scaled chip.
Finally, the user can change the number of transistors on the
reference chip and study the dependence of the various chip
performance indicators on its integration level. In addition
to displaying the simulation results on the screen, the
program also writes the various chip parameters and the
corresponding results for the reference chip, scaled chips and
the chips with different integration levels on an output file
called "NCHIPSIM.OUT".
95-11
V. SIMULATION RESULTS:
The program "NCHIPSIM” ha^ been used to compare the
simulation results for several actual NMOS microprocessor
chips to the available data [5] . For example, such a
comparison for the 1.5-micron NMOS microprocessor chip called
HP Focus (1982) is shown in Table 1 and that for the 3-micron
NMOS microprocessor chip called Stanford MIPS (1984) is shown
in Table 2. The program has also been used to study the
dependence of the chip performance on its minimum feature size
as well its integration level i.e. the number of transistors
or gates on the chip. For example, Figures 1 and 2 show the
dependences of the maximum cloclc frequency and the power
consumption of an NMOS microprocessor chip on its minimum
feature size in the range 0.1-2. 5 microns while Figures 3 and
4 show the maximum clock frequency and the power efficiency
for a 1 -micron NMOS microprocessor chip on the number of
transistors on it in the range 100-100,000,000 respectively.
VI. RECOMMENDATIONS:
At present, the program "NCHIPSIM" is expected to
simulate NMOS integrated circuit chips of feature size 1
micron or greater quite accurately. However, for simulating
the submicron chips correctly, the algorithm used in NCHIPSIM
must be modified to include the high-density effects.
95-12
Thereafter, this work must be extended and the
appropriate mathematical algorithms and the related software
modules must be developed for the integrated circuit chips
based on the silicon CMOS, silicon BJT, GaAs FET, GaAs HBT and
the other state-of-the-art VHSIC technologies. It is important
that, for the small-geometry high-density Chips, the
capacitance calculation routines include the contributions of
the fringing fields, coupling capacitances with the neighbors
and the effects of shielding by the neighboring conductors.
For the high-frequency GaAs chips, the interconnection
resistance calculations should consider the skin effect, the
gate delay calculations should include the longitudinal as
well as the transverse delays in the field effect transistors
and the interconnection delay calculation., should include the
effects of the distributed capacitances and inductances in the
on-chip interconnections.
It was decided that the author will continue with the
summer research effort and carry out the research work
recommended above at the author's institution with funding
from the Research Initiation Program of the Air Force Office
Of Scientific Research.
95-13
REFERENCES
1. Donath, W.E., "Placements and Average Interconnection
Lengths of Computer Logic, " IEEE Trans. Circuits and Systems,
Vol. CAS-26, April 1979, pp. 272-277.
2. Landman, B.S. and R.L. Russo, "On a Pin versus Block
Relationship for Partitions of Logic Graphs, " IEEE Trans .
Computers, Vol. C-20, Dec. 1971, pp. 1469-1479.
3. Price, J.E., "A New Look at Yield of Integrated Circuits,"
Proceedings of IEEE. Vol. 58, Aug. 1970, pp. 1290-1291
4. Bakoglu, H.B. and J.D. Meindl, "Optimal Interconnection
Circuits for VLSI," IEEE Trans. Electron Devices. Vol. ED-32,
No. 5, May 1985, pp. 903-909.
5. Toong, H.D. and A. Gupta, "An Architectural Comparison of
Contemporary 16-Bit Microprocessors," IEEE Micro. Feb. 1983,
pp. 26-38.
95-14
Maximum Clock Frequency (MHz)
in
cJ
CM
ol
« w
<D Q.
Si
O O)
jg .5
^ w
CO 3
S u •
DOM
SMC
•H M O
X <U M
(TJ O U
S 0-H
^ 2
0) a
x: o
U >4 •
° s H
O g»,o
r ^ 2 Q)
•- ^ I ^
2 c
C (U
M s
O4 o
0) o -u 0)
M "CJ M J3
tN .H JJ
J- A <y M
UJ ^ C c
4J nj -H
w ij <1>
"S Og ”
0)
LL 0° D
•H ° ij
Etjin ^
3 M M 14^
Sod
?= H s
= ^ >1-^
2 M
2 g.M
Power Consumption (W)
m
(N
u a
0) -H
3 ^
o o
0)
x;
w
c
o
o
to
4J to O
(1) -H
M-l U 6
O O
IT)
0) a •
O O CN
C ^ I
0) O IH
"0 -H •
CEO
0)
CuV) 4)
0) O 01
V s
z
<D
x:
iJ
c
(0
o
V jC
IXI to 4J
•H
to
c
(TJ
c
o
•H
0)
c
•H
V 4J
o
a;
N
•H
W
0
•H O
•o o
0) o 0)
Cuo
O 4J
S CN (tJ
M 4)
t/5 (T5 4-1
M 4^
z o
o
z
CN
31
c
o
'H
4J
a
s to
3 XJ
to -rH
01 c
o c:
u o
95-16
Maximum Clock Frequency (MHz)
95-17
100 1,000 10,000 100,000 1,000,000 10,000,000 1.000E+08
Number of Transistors
Figure 3: NCHIPSIM prediction of the dependence of the maximum
clock frequency of a 1-micron NMOS microprocessor chip on the
number of transistors on it in the range 100-100,000,000.
r Efficiency (GHz/mW)
00
o
+
111
o
iZ O
- O
O
(0
w _
u ^
S CO
o ^
g o
L O
- O
= O
‘I o'
!- O
• o
o
h cS (/)
r S O
o5
’w
c
“ o Ci3
_ 9. H*
- 2
- o o
00 o>
9--
is
a c o
0) ° O
s:
jj ao
•H o
M-i x: o
o o s
o
a) j. o
OS’-*
G m I
0) S o
TD V, O
C M H
o
a 2
0) cn
”0 q" c
o c
G S -H
O Z
o
0
C
n —
•H
t3
M 0
o 1
o ^
0)
iH
a
u ^
s
1
9 Z
s
1 0
iJ
o
M
CO
CO
nj -H
Oj
V)
H
X
oj G
o ^
u
° (-1
z
4J
• •
O <«
o
*3*
C 0
o
0)
<D
•rH U
S-l
o d)
T**
3
•H Xi
a
^ e
•r-}
4-1 3
Cu
<i> c
TABLE 1 : Actual Data and Simulation Results for
NMOS Microprocessor Chip HP FOCUS (1982)
95-19
95-20
1990 USAF-UES SUMMER FACULTY RESEARCH PROGRAM/
GRADUATE STUDENT RESEARCH PROGRAM
Sponsored by the
AIR FORCE OFFICE OF SCIENTIFIC RESEARCH
Conducted by the
Universal energy systems, Inc.
FINAL REPORT
Application of Photoreflectance to Novel Materials
Prepared by:
Muhammad Z. Numan
Academic rank:
Assistant Professor
Department and
Physics Department
University:
Indiana University of
Pennsylvania
Research location:
USAFAVRDC
Wright Patterson A.F.B.
Dayton, OH 45433
USAF Researcher:
M. Omar Manasreh
Date:
Sept 15, 1990
Contract No:
F49620-88-C-0053
Application of Photoreflectance to Novel Materials
by
Muhammad Z. Numan
ABSTRACT
Photoreflectance spectroscopy was applied to the InGaAs/GaAs single
quantum well structures of different well thickness and to the low
temperature molecular beam epitaxy grown GaAs cap layers on both n-
and p- t^'pe GaAs substrates. The PR spectra at both room temperature
and 77K have been studied. The GaAs study clearly indicates a lowering of
surface potential associated with the unpinning of the Fermi levels
reported for these systems. Both 200®C and 400'’C caps demonstrate a
disappearence of the Franz-Keldysh oscillation. Suggestions for future
experiments are made.
96-2
Acknowledgements
I wish to thank the Air Force Systems Command and the Air Force Office of
Scientific Research for sponsorship of this research. Universal Energy
System has made the summer very comfortable by their careful attention
to the needs of the participents.
The research experience was truly enjoyable due to the cordial interaction
and constant encouragement from Dr Omar Manasreh, who supplied most
of the samples for the work on the single quantum wells. Dr. David Look has
provided samples and support for the work on low temperature MBE GaAs.
John Hoelscher has worked closely with me lending freely of his technical
expertice. Special thanks are due to Dr. Phil Yu, Dr. Keu Evans, Capt
Johnston, and Gail Brown for their cooperation. The encouragement and
guidance of Col. K. Soda has been invaluable throughout the project.
96-3
L INTRODUCTION:
Photoreflectance (PR) spectroscopy is becoming an important tool for the
characterization of bulk semiconductors and semiconductor
microstructures, such as quantum wells and superlattices 1. Its popularity
stems from the nondestructive, contactless, and room temperature nature
of the technique, involving no special sample preparation or mounting
requirements. PR found notable application in.measuring direct gap of
compound semiconductors, intersubband transitions in multiple quantum
wells (GaAs-AlGaAs, GaAs-InGaAs), and change in surface Fermi level of
GaAs caused by photowashing2.
At the WRDC/ELRA laboratory there exists excellent facilities for both room
temperature and low temperature PR. In the past, summer fellows have
used these facilities in fruitful collaboration with resident scientists^. Two of
the areas of interest to the laboratories, namely, GaAs-InGaAs single
quantum wells and low temperature MBE grown GaAs passivation layers,
lend themselves to PR analysis with prospects of unique and important
results. Current research have been undertaken to explore these
possibilities.
My research interest in the past has spanned various experimental aspects
of semiconductor materials and devices including structural (Rutherford
backscattering and channeling) and electrical characterization. Recently, at
lUP Physics department, we developed an optical characterization
laboratory with funding from Research Corporation and local grants. A
vacuum far infra-red FTIR system (Bio-rad/ Digilab FTS-40V) has been
added to the lab. The summer fellowship will strengthen our program at
lUP and establish fruitful collaboration between the labs in the areas of
materials characterization and analysis.
96-4
n. OBJECnVES OF THE RESEARCH EFFORT:
This summer project addresses two different materials systems of
technological interest. The objectives in each area are delineated below.
Firstly, we consider strained quantum wells of luxOai.^As in GaAs for room
temperature PR study of conduction to valence band transitions of
quantized energy levels. Strain dependent calculation of the transition
energies as a function of well width can be compared with the experimental
peaks in PR spectra to extract band offset for this system. We decided to
grow MBE samples of InxGai-xAs/GaAs single quantum wells (x=.10 - .20)
of well width ranging from 90A - 150A for PR measurements at room
temperature and 70K. The results will be fitted with existing and new
theoretical models to extract band offset parameter
Secondly, PR will be used to understand the reduced surface pinning of low
temperature MBE grown GaAs layers. According to recent reports, the
well known surface pinning effect of the Fermi level in GaAs is reduced by
75% when a low temperature MBE grown layer is used to cap the doped
GaAs^. Photoreflectance is a sensitive tool in determining surface electric
field in bulk materials^. For large fields, Franz-Keldysh oscillations appear
in the PR spectra in contrast to the low field case^. If the surface field in
the capped GaAs is significantly lowered a shift in the period of the FK
oscillation will permit us to measure the change from room temperature
PR spectra. Samples of both p-type and n-type MBE GaAs will be capped
with 200A layers of low temperature material. PR spectra will be
compared with control samples to see if a change in FK oscillation can be
observed.
m. PHOTOREFLECTANCEDATAANALYSIS:
In Photoreflectance spectroscopy, a chopped laser beam is used to
modulate the built in electric field at the surface or an interface of the
semiconductor (Fig. 1). The relative change in reflectance is then recorded
96-5
using lock in detecton method. Photoreflectance spectra is customarily
analyzed by the third derivative functional form due to Aspenes^,
according to:
AR/R = Re[C e‘9(E - Eg +ir)-n ] ( 1 )
where E is the energy of the probe beam, G is a broadening parameter for
the critical point energy E, and n refers to the type of critical point. Eqn (1)
applies to the case of low built-in electric field typified by the absence of
Franz-Keldysh oscillation. In the presence of FK oscillation, AR/R takes an
asymptotic form:
AR/R = Cos {j[(E - Eg )/ li0]3/2 + (}>} (2)
where he is a characteristic energy,
( 110)3 « e2li2 F2 /2p. (3)
Photon energy at the m-th extrema, Em , satisfies:
mp = j[(Em - Eg )/ Be ] 3/2 + <j, (4)
4
A straight line fit of ^[(Em - Eg )]3/2 vs. m gives the characteristic energy,
he. Surface electric field, Fs, can then be found from Eqn. (3).
rv. RESULTS:
Fig. (2)-(3) shows the room temperature and 80K photo-reflectance data
from the single quantum wells of well width 90A and 140A. We have not
been successful in analyzing this data using TDFF method. A more
comprehensive fitting procedure is being developed.
96-6
Fig* (4)-(5) shows the results for the low temperature capped layers
overlayed with the uncapped samples. The presence of FK oscillation is
evident in case of the uncapped (control) samples, which disappears for the
capped layers grown both at 200‘’C and 400®C. This result is a direct
evidence of the reduction in band bending for the latter cases. In Fig. (5)
G2-994 is the uncapped Be doped sample and G2-994 is the capped
sample. G2-992 is the n-type uncapped sample of Fig. (4) for comparison
of FK oscillation period in the two impurity types. In the p-type GaAs an
additional low energy feature was noticed, which may be attributed to the
shallow impurity levels.
V RECOMMENDATIONS:
The photoreflectance results for the low temperature MBE cap layers are
very interesting and could resolve many unanswered questions about this
technologically important material. Low temperature measurements
should be performed on these samples to see if FK oscillation can be
observed in the capped material. This will allow one to determine the build
in field and the surface potential quantitatively. The low energy feature
should be studied more carefully to ascertain its origin.
As to the quantum wells, the signals are not well resolved in the room
temperature spectra. The MBE chamber was known to have some problem
when these samples were grown. New samples should be grown carefully
and the measurements repeated. To improve the apparatus, a new grating
should be used in the monochromator of the room temperature system
along with a Ge detector to sufficiantly extend the wavelength range to
allow one to study higher mole fraction of Indium in the well. A variable
neutral density filter can be placed in front of the monochromator as
shown in Fig. (6) to have uniform intensity of the probe beam. A follow-up
mini grant proposal will be submitted to achieve some of these
recommendations.
96-7
REFERENCES
1. N. Botka, D.K. Gaskill, R. S. Sillmon, R. Henry and R. Glosser, "Modulation
Sjpectroscopy as a Tool for Electronic Materials Characterization", J. Electr
Materials, U. ( 161),1988.
2. F. H. Poliak, and H. Shen, "Photoreflectance Characterization of
Semiconductors and Semiconductor Heterostructures", J. Electr Materials,
11 (399),1990.
3. Michael Sydor, J. R. Engholm, M. 0. Manasreh, C. E. Stutz, L. Liou, and K. R.
Evans, "Photoreflectance and the electric fields in GaAs depletion region"
Appl Phys. Lett. 5^.(1 769), 1990.
4. D. C. Look, C. E. Stutz, and K. R. Evans, "Unpinning of GaAs Fermi Level by
200'’C Molecular Beam Epitaxial Layer", Unpublished.
5. R. N. Bhattacharya, H. Sen, P. Parayanthal, and Fred H. Poliak,
"Electroreflectance and Photo-reflectance study of the space charge region
in semiconductors: (In-Sn-0)/InP as a model system", Phys. Rev.
B37(4044)1988
6. Fred H. Poliak and 0. J. Glembocki, "Modulation Spectroscopy of
Semiconductor microstrlicrures: an Overview", SPIE Proceedings, Vol. 946,
P. 2, 1988.
7. D. E. Aspenes, "Third Derivative Modulation Spectroscopy with low-field
electroreflentance". Surface Science 37418(1973)
96-8
96-10
LOW TEMP MBE
P typo material
siiNfi eav
nanometers
G2-993 - control (GZ--992) - G2-994
LAMP
^NOCHROMATOR
SAMPLE
AR/R
LASER (OR OTHER "■
SECOHOARY LIOHT SOURCE)
96-14
1990 USAF-UES SUMMER FACULTY RESEARCH PROGRAM
Sponsored by the
AIR FORCE OFFICE OF SCIENTIFIC RESEARCH
Conducted by the
Universal Energy Systems, Inc.
FiNALBSEOBI
Electfonic^tructure and deep impurity levels in GaAs related
compound semiconductors and superlattices
Prepared by:
Academic Rank:
Department and
University
Research Location:
USAF Researcher
Date:
Co.ntract No:
Devki N. Talwar and Alan Coleman
Assistant Professor and Graduate Student
Department of Physics
Indiana University of Pennsylvania
Electronic Technology Laboratory
ELRA/WRDC, Wright Patterson Air
Force Base, Ohio, 45433
Dr. M. 0. Manasreh
Sept. 27, 1990
F49620-88-C-0053
Electronic structure and deep impurity levels in GaAs related
CQmpQupd semicQn(j.uctQrs and .superlattiggg
by
Devki N. Talwar
Alan Coleman
ABSIBAGI
The band structure of periodic, ultra-thin, iattice matched (AlxGai.xAs)^
/(GaAs)n, and strained layer (GaAs)m/{lnxGai.xAs)n superlattices (SL's) grown
along the three main orystallographic orientations (001), (110), and (111) is
studied by using a second-neighbor tight binding theory. The SL wave functions
are described as a linear combinations of bulk Bloch functions {sixteen, if spin is
included) for each of the two constituent materials while the alloy Alx Gai-xAs (or
Inx Gai-xAs) is treated in the virtual crystal approximation. To incorporate the
effects of strain in strained layer SL’s, a new method is developed, based on
Harrison's scaling scheme which properly includes the variation of bond
lengths and bond angles. While studying the band structure, we found that the
band gap in SL's depends not only on the layer thicknesses through quantum
mechanical effects but also through the strains in the constituent layers. Our
calculated results for the band structure of (GaAs)i/(AIAs)i and (GaAs)2/(AIAs)2
SL’s are found in excellent agreement when compared with the existing
sophisticated self-consistent pseudopotential data. Using the above information
of band structure, a Green's function theory of impurity levels is being
developed. This theory will allow us to predict the role of deep levels in SL's
and QW’s and may prescribe a method to overcome the effects of deep traps
which we believe are responsible for limiting the performance of Si-doped
HEMT’s, and other technologically important devices.
97-2
Acknowledgements
The author wishes to express his sincere thanks to the Air Force System
Command (AFSC), and the Air Force Office of Scientific Research (AFOSR) for
the award of a "1 990 UES Summer Faculty Research Fellowship" that enabled
him to work at the Electronic Technology Laboratory, Wright Patterson Air Force,
Base, Ohio. Thanks are due to Dr. Rodney C. Darrah, of UES and to Col.
Claude Cavender for arranging an excellent slide show which provided the
participants an indepth knowledge of the research intersts of the AFOSR.
The ten week stay at WPAFB was short but very exciting, enjoyable, and
rewarding. Col. Ken Soda and Dr. Omar Manasreh of the Electronic technology
laboratory (ELR/ ' WRDC) provided Ihe author with support, encouragement,
and a truly friendly working atmosphere. Fruitful discussions on the subject
matter with Drs. Dave Look, Omar Manasreh, Phil Won Yu, D. Reynolds, Keith
Evans, Ed Stutz, Dave Fischer, Dennis Whitson, and Cole Litton were
invaluable in learning the new and exciting area of semiconducting
superlattices.
97-3
I. INTRODUCTION
Modern crystal-growth technique such as molecular-beam epitaxy (MBE) has
made possible the laboratory synthesis of a variety of ultra-thin layer
superstructures. Among others, there is a special group of systems which
spontaneously construct the superstructure under certain experimental
conditions. One of them is the monolayer based strained GaP/lnP structure.^-s
Other systems include lattice matched GaAs/AIGaAs, and strained GaAs/InGaAs
quantum wells (QW's), and superlattices (SL’s). The strained GaP/lnP structure
exhibits an atomic ordering which has been revealed recently by the transmission
electron microscopy to be a (111) SL Strained layer materials exhibit special
importance as they permit study of the relation between band structure and strain.
Again they retain all the advantages of lattice-matched systems and exhibit
additional features, such as the potential for the highest valence states to be light-
hole-like, which Is of benefits for many tailorable engineering applications.
A major advance for potential high performance devices was made when
Stromer, et al.^ introduced n-type doped modulation doped samples. The
underlying idea is that, at equilibrium, charge transfer occurs across a
heterojunction to equalize the chemical potential (i.e., the Fermi level) on both
sides. If the wide band gap side of AIGaAs/GaAs heterojunction is doped, the
electrons will be transfered to the GaAs layer until an equilibrium is reached. This
occurs because electron transfer raises the Fermi energy on the GaAs side due to
level filling and aiso raises the electrostatic potential of the interface region
because of the more numerous ionized donors in the AIGaAs side. The charge
tansfer effect makes possible an old dream of semiconductor technologists, i.e.
getting conducting electrons in a high-purity, high-mobiiity semiconductor without
97-4
having to introduce mobility-limiting donor impurities. Since then, modulation
doping has been applied to a number of situations involving various
semiconductor pairs and also to hole modulation doping. In recent years, the
impressive development of the subject has resulted in several high-performance
electronic devices including faster complementary logic circuits,^ efficient and
stable long-wavelength lasers,^ and high electron mobility transistors (HEMPs),
etc.^®’^2
II. Need for a reliable theory of band structure in SL's
a. The Present Situation
In device design using heterostructures, it is imperative to know the electronic
band structure and the behavior of defects in SL’s and QW's from a reliable
theoretical model. In bulk lll-V semiconductors, the point defects have already
played a leading role: they compensate and scatter the free carriers, shorten their
life time, and introduce spurious effects. This is the case, for instance, of the
famous EL2 defect in GaAs and OX center in AlxGa^.xAs.’’^ in heterostructures, in
addition to point defects located in the layers, there may also be the possibility of
the existence of interface defects between layers.''^ Thus, an understanding of
the role of defects in heterostructufes is atleast as important as in bulk or ternary
compound semiconductors. However, very few studies have been performed in
order to characterize defects and to predict their behavior in SL's.
b. Objectives
Eariierwe developedis-te an empiricai second-neighbor tight-binding theory
and applied it successfully to treat the energy-band structure (See Figs. 1 a-c)and
97-5
deep impurity levels (See Figs. 2 a-b) in elemental and IllrV compound
semiconductors. The strength of our approach was in its chemical trends of the
matrix elements of the TB Hamiltonians, since these are the local functions of the
bulk semiconductor. The observed trends'^s-ie jn the short-range interactions are
particularly suited for the study of the interface disorder and for the inclusion of
strain effects in strained-layer SL's. To incorporate strain in strained-layer SL’s
we proposed here a new method, based on Harrison’s scalirigi^ scheme, which
includes the variation of the bond lengths and bond angles. With these ideas
properly considered, a systematic transition of the TB method from bulkf5-i6 to
SL’s is possible for studying both the band structure and the impurity levels. Such
a theory is needed for predicting the conditions under which a specific impurity
produces a shallow donor level and "dopes” the semiconductor, versus the
possibilities of deep-level formation and the trapping of charge carriers. The
present work done under 1990 LIES Summer Faculty Research Program (SFRP)
had major objectives; (i) to develop a second-neighbor Tight-binding theory
for the Band structure of SL’s, and (ii) to formulate the TB-Green’s function theory
for deep-levels in SL’s
III. Theory
a. Band structure
The method used for the caiculations of the band structure of SL is similar to
that of Ihm^s with two exceptions: First, formulae for the TB matrix elements have
been corrected to yield the same output for the bulk X-point energies as the input
data. Second, the interactions up to and including second nearest neighbors
97-6
have been introduced to allow an exact description of L-point energies in bulk
semiconductors.
Consider an SL of two different alternating zinc-blende crystals labelled ‘ab’
and ‘AB’ with (001) interface, where a(A) and b(B) are cation and anion atoms,
respectively. The SL (ab)n/ {AB)^ contains 2(n+m) atoms in a unit cell at R| with
four (sp^ orbitals each ; |ay> where a denotes the s, x (=px), y (=Py) and z (»pz)
orbitals, and j represents the site index in a unit cell which runs from 1 through
2(n+m). If the Bloch sums defined by
i
are used as basis, the SL Hamiltonian can be written as
97-7
where k = “(kiika.ka) (a© is the lattice constant) is the wave vector, xj is the atom
position in a unit cell and N is the number of Braivais lattice points. Each element
in Eq. (2) represents a [4x4 (or 8x8, if spin is included)] matrix and, for instance,
ab {ba) and aa {bb) correspond to nearest neighbor interactions {<Xj“lH°|Xj+iP>}
and second- nearest neighbor interactions {<Xj“|HO|xj+2M having j of site 'a' ('b
'). The diagonal elements as a or b symbolize {< Xj“|H°IXjP>} with j of site 'a 'or
'b' and contain both intrasite energies and second neighbor interactions.
Elements bA, Ba, aA and bB correspond to intermaterial interactions at interface.
Additionally, the Hamiltonian for the bulk crystal ab is expressed as follows:
Ho (na1 , m=0) «
'a + aa + aat ab + bat
iba + abt c+cc+oct
(3)
which is equivalent to that reported by Talwar and Ting.'' 5 The TB-parameters
obtained earlier'5-i6 for the bulk lll-V compounds are assumed not to change at
the interior layers of the corresponding compound when the constituent materials
are combined to form a SL. In addition to the bulk constituents, the interface
region must be described properly. For interface region, we will use averages of
the bulk parameters for both compounds or the atoms are assigned to either bulk
regions. A properly specified valence-band offset, combined with the bulk
electronic structure, will determine the rest of the band line-ups.
b. Inclusion of Strain Effects
A new method is suggested here based on Harrison’s scaling scheme'^, to
incorporate the effects of strain in the band structure calculations of SL's. Unlike
97-8
other methods this method does not require experimental data for the shifts or the
splittings of energy bands and has the advantage of being applied to any strained
system. Two extreme cases are considered here. One of them is the so-called
free standing model, which describes the strain distribution when the thickness of
the SL exceeds a certain critical value or the SL is lattice matched to the
substrate. The other case considered here is the coherent growth model which
has been suggested to be true both experimentally and theoretically when the
thickness of the SL is under the critical value.
c. TB-Green's Function Theory of Deep-Levels in SL's
Quite recently in random alloys such as AlxGa^.xAs, we have proposed^d the
possibility of "shallow-deep” transition for isolated donor defects (DX-center),
when X reaches a critical value. It is our belief that the deep levels are much less
sensitive to changes of alloy composition x than the band edges. Slightly
overstating the point, we may also say that the deep levels are almost constant in
absolute energy, and that by varying x one can cause the conduction or valence
band edge to move through the deep levels. In AI^Ga^xAs, for example, by
changing x the band gap varies and it cpens up a window of observability of the
deep levels (c. f. Fig. 2). Thus by a proper selection of the alloy composition x in
mixed semiconductors, a given impurity can produce a shallow level rather than a
deep one, is a form of "band-gap engineering". We believe that similar ideas of
"band-gap engineering" can be achieved by altering the atomic ordering in the
properly selected semiconductors which make the hosts either nearly lattice
matched or strained-layer SL's, rather than a random alloy.
97-9
IV. Preliminary Results
The tight-binding method outlined above (c.f. Sec. II) offers a useful frame-work
for the subband energy calculations in both lattice matched and strained-layer
SL's. Here, we present via our preliminary calculations, a picture of the
successful transition of the bulk TB-theory to the band structure calculation of
GaAs/AlxGai.xAs (001)SL's. By comparison with the existing sophisticated
theoretica|20-2i (c.f. Fig. 3) and reliable optical (PL or PLE) data,22-23 yve have
obtained an extensive knowledge which supported our IB approach. While the
methodology adopted and the group theory used are of general applicability, the
choice of one example presented here will leave out many important issues that
we are planning to undertake. These issues are:
a. Calculation of Energy Band Gap variation with period of the SL’s
This type of study may help us attain new band gap regimes not otherwise
achievable with the bulk lll-V compound semiconductors. For instance, InAsSb
strained layer detectors are believed to operate at wavelengths up to about
'“12pm and may possibly be more stable than HgCdTe detectors.
b. Cross over from Type-i to Type-ll (GaAs)ni/(AiAs)ni SL's
An equally important and still unresolved issue in short-period
(GaAs)^(AIAs);jj SL's is to understand the cross over (c.f. Fig. 4) from type-l to
type-ll superstructure. Here, m represents the number of monolayers in one SL
period (a monolayer of GaAs or AlAs is ~'2.83 A thick). In Fig. (4a), we have
displayed both T-like quantized states (which are mai ily confined in GaAs layers)
97-10
and X-like states (which are mainly confined in AlAs layers) |n (GaAs)n,/(AIA£f)n
SL's. When m is relatively large, the r-like energy levels are lower than the X-like
levels and the resulting superlattices are of type-1 in which both electrons and
holes are confined in the same constituent layers (see Fig. 4a). However, when m
is small enough, the X-like states will become the lowest states in the conduction
band. The superlattices become type-ll in which the electrons and holes are
confined in the adjacent constituent layers, respectively, (as shown in Fig. (4b)).
Obviously, for certain critical values of m the r-like levels are nearly equal to the
X-like ones Le. a type-1 to type-ll transition occurs as shown in Fig. (4c).
Experimental evidence for such a crossover is likely to be important^^ for the
vertical transport (tunnelling) in SL structures. Although, some PL data is known
for the critical thickness (m) for which the transition from type-1 to type-ll occurs at
ambient pressure. However, the existing results are rather sparse primarily
because in type-1 SLs the X-like states are difficult to detect either by PL or by
PLE measurements.22*23
c. Relative Position of the X-vaileys in Type-ll (GaAs)m/(AIAs)m SL’s
Another important and unsolved problem, from the basic physics standpoint, is
to know which X-valley states (in the conduction band), the Xxy and or the Xz in
type-ll short-period (GaAs)^(AIAs)^ SL’s attain the minimum energy with respect
to the valence band maximum. Here, Xz is the X-valley with wave vector (k)
parallel to the growth axis, while Xx and Xy have k in the plane of the layer. From
simple Group-theoretic analyses and by considering appropriate Brillouin zones
of the SL’s, it can be shown that for the (001 ) SL's, there is a possibility of mixing
of r-Xz states. On the other hand, for the (01 1 ) SL, it is the Xz - Xx states that mix
97-11
together and plays a dominant role, while for the (111) SL, it is the mixing of L-
valley with the r-valley state which is important. In SL's the mixing and splitting of
the energy states is manifestly an interface effect which cannot be treated within
the Kronig-Penney or k.p (EMA) models. This issue, we believe can be resolved
only if realistic theories are considered and if the band structure of
(GaAs)m^(AIAs)^ (Oil) short-period SL’s is calculated.
d. Problems Related to the Modulation Doping in AIGaAs/GaAs SL's
Modulation doping in GaAs/AIGaAs SL's, by which donor (say. Si) impurities
are inserted into the large gap AIGaAs layers of a SL but donate their electrons to
the small band-gap GaAs layers, has already played a role in the development of
high speed lll-V compound semiconductor devices. However, practical devices
based on AI^Ga^.^As are often limited to alloy compositions x < 0.3. This is due to
the formation of Si-related defects that are deep traps. Some devices, such as
H£MT’s operate using QW structures at or near GaAs/AIGaAs interfaces and the
performance of these devices depends on the donor-doping, the alloy
ccmposition and the SL’s structure. Clearly, the tight-binding Green's function
theory in SL’s and heterostructures will help us learn the electronic
characteristics of point defects giving rise to deep-levels.
97-12
References
1. Smith, D. L, andC. Mailhiot, Rev. Mod. Phys. 1990, Vol. 62, p. 1.
2. O'Reilly, E. P. Semiconductor Sci. Tech. 1989, Vol. 4, p, 121.
3. Hayakawa, T. T., K. Suyama, M. Takahashi, M. Kondo, S. Yamamoto and
T. Hijikata, Appl. Phys. Lett. 1988, Vol. 52, p. 339.
4. Dingle, R., Editor 'Applications of Multi-quantum wells, Selective Doping, and
Superlattices' 1987, (Academic, New York).
5. Osbourn, G. C., R. M. Biefield, and P. L. Gourley, Appl. Phys. Lett. 1982, Vol.
41, p. 172.
6. Fritz, I. J., L. R. Dawson, and T. E. Zipperian, Appl. Phys. Lett. 1983, Vol. 43,
p. 846.
7. Stdrmer, H. L., R. Dingle, A. C. Gosard, W. Wiegmann, and R. A. Logan, Conf.
Sen -Inst. Phys. 1979, Vol, 43, p. 557.
8. Lepcre, A. N., M. Levy, H. Lee, and E. Kohn, Electron Lett. 1988, Vol. 24,
p.364.
9. Henderson, T. et al., IEEE Electron Dev. Lett. 1986, Vol. EDL7, p. 649.
10. Mzuta, H. et al., IEEE Trans. Electron Dev. 1989, Vol. 36, p. 2307.
11. Watanabe, Y. et. al., IEEE GaAs 1C Symp. Dig. Tech. Papers 1988, p. 86.
12. Notomi, S. et, al., IEEE GaAs 1C Symp. Dig. Tech. Papers 1987, p. 177.
13. Henning, J. C. M., and J. P. M. Ansems, Semicond. Sci. Tech. 1987,
Vol. 2, p. 1 .
14. Bourgoin, J. C., and H. J. von Bardeleben (private communication)
15. Talwar, D. N., and C. S. Ting, Phys. Rev. 1982, Vol. B25, p. 2660.
16. Talwar, D. N., K. S. Suh, and C. S. Ting, Phil. Mag. 1986, Vol. B54, p. 93 :
K. S. Suh, Ph.D. Thesis (University of Houston, 1988), (unpublished).
97-13
17. Harrison, W, A. in Electronic Structure and the Properties of Solids,
1980 (Freeman, San Francisco).
18. Ihm, J., Appi. Phys. Lett., 1987, Vol. 50, p. 1068.
19. Talwar, D. N., M. 0. Manasreh, K. S. Suh and B. C. Covington, Appi. Phys.
Lett. 1987, Vol. 51, p. 1358.
20. Gilbert, T. G., and S. J. Gurman, Superiattices and Microstructures, ‘\9Q7,
Vol.3, p. 17.
21 . Srivastava, G. P. in Gallium Arsenide and Related Compounds 1987 (Inst.
Phys. Conf. Ser. 91) p. 529.
22. Jiang, D; S., K. Kelting, T. Isu, H. J. Queisserand K. Ploog. J. Appi. Phys.
1988, Vol. 63, p. 845 .
23. NaKazawa, T., H. Fujimoto, K. Imanishi, K. Taniguchi, C. Hamaguchi,
S. Hiyamizu, and S. Sasa, J. Phys. Soc. Jpn. 1989, Vol. 58, p. 2192.
24. C. Minot, H. Le Person, J. F. Palmier and R. Planel, Superiattices and
Microstructures 1989, Vol. 6, p. 309.
97-14
Figure captions
Fig. 1a. Calculated band structure of GaAs based on the second-neighbor tight-
binding parameters of Talwar and Ting (Ref. [29]).
Fig. 1 b. Same key as of 1 a but for AlAs.
Fig 1c. Calculated band structure of Alo.43Gao.57As in the virtual-crystal
approximation.
Fig. 2a. Schematic illustration of the dependence of deep impurity levels,
caused by donor defects (DX-centers) occupying cation sites, on x in
AlxGal -xAs (after Ref. [1 9]). The zero of energy is the valence band
maximum of the alloy.
Fig. 2b. Same key as of 2a but for donor defects on anion sites.
Fig. 3. Calculated electronic band structure of (001 ) oriented (GaAs)i/(AIAs)>|
superlattice based on our second-neighbor tight-binding theory
described in the text. The results are found in excellent agreement with
the sophisticated seif consistent pseudopotential calculation.
Fig. 4 Schematic illustration of the energy band structure for the
(GaAs)fr/(AIAs)fjj SL's of the transition (from the r-like state to the
heavy hole state) and the transition (from the X-like state to the heavy
hole state). The dashed lines represent the X-valley energy of GaAs and
AlAs bulk materials. The bold lines represent the lowest quantized
energy states in the wells, (a) The type I SL. (b) The type ii SL. (c) The
SL at the transition point from the type- 1, to type- II SL. (a') Case (a)
under hydrostatic pressure, having been the type II SL.
97-15
6
Fig. 3
>
4
2
0
>
Ui
>
G
(£
UJ
Z
UJ
-2
4
■6
•8
-10
-12
-14
-
* • * * * »
* ♦
' (
- • .
*’** .
**'*•»». . ! I • *
* • . . . . . * • *
1 1 1 • • » 1 » # ♦ 1 J J * , , ,
1
GaAs/AIAs
•
SLS (OOI)’-'...
* •
1 « •
-
_ L..1 _ 1— 1
' ! • *
L-i _ 1 _ 1 _ 1 _ _ _ 1 _ 1 _ III 1.1. 1 — 1—
0 5 10 15 20253035404&50
WAVE VECTOR
97-18
1990 USAF-UES SUMMER FACULTY RESEARCH PROGRAM
GRADUATE STUDENT RESEARCH PROGRAM
Sponsored by the
AIR FORCE OFFICE OF SCIENTIFIC RESEARCH
Conducted by the
Universal Energy Systems, Inc.
FINAL REPORT
sues 1X1
iirnarouiid
Prepared by: John S. Bay, Ph.D.
Academic Rank: Assistant Professor
Department and University: Bradley Department of Electrical Engineering
Virginia Polytechnic Institute and State University
Research Location: WRDC/FIVMB
VVPAFB, OH 45433
USAF Researcher: Mangal D. Chawla, Ph.D.
Date: August 1, 1990
Contract No.:
F49620-88-C-0053
Sensor Integration Issues in Robotic Rapid Aircraft Turnaround
by
John S. Bay
A-B§IBAgI
Concepts for robotic rapid aircraft turnaround arc examined for the anticipated sensor
specification and data processing requirements. Of particular concern are the inspection, monitoring,
and supervision of robotic operations in an integrated combat turn where there is the possibility of
chemical and biological hazards. Because of the harsh environment expected, it is suggested that the
robotic refueling operations exploit infrared sensing systems, including thermal imaging and range
sensing, as opposed to ultrasonic measurement techniques. Parts mating operations should also
incorporate active and passive compliance devices, as well as force and torque sensing. For some
operations, such as ordnance loading, it is suggested that force>amplifying telemanipulators be
considered over full autonomy. Also, some sensing apparatus should not be mounted to the robotic
components themselves, but are more appropriately fixed in space. These fixed sensors should
include vision and floor-mounted load cells. For laboratory feasibility studies and demonstration,
tactile sensor research is suggested, along with a data processing technique based on distributed
computation of the extended Kalman filter.
98-2
I. INTRODUCTION
A. The Problem
During wartime, a critical consideration in the generation of sorties for tactical fighter
aircraft is the time required for ground servicing. An integrated combat turnaround (ICT). is the
procedure used by the U. S. Air Force for maximizing this aircraft turnaround rate. During this
procedure, the aircraft must be marshalled to the servicing area, the wheels must be chocked, the
aircraft must be inspected for any damage, and electrically grounded and bonded to the refusing
vehicle. Following a well specified set of rules, the aircraft is then refueled and re-armed with
missiles and bombs before it - returned to the runway for the next mission.
Of additional concern in the operation is the possibility of chemical and biological (CB)
warfare hazards on the flight line. In such a situation, all equipment and manpower must be
protected from these CB agents, and must be decontaminated subsequent to exposure.
Unfortunately, the protection gear worn by the ground crew consists of several layers of bulky
material which severely restricts their dexterity and endurance. Ideally, the turnaround is targeted
for 45 min. servicing time, but the crew members can suffer heat stress during this time in the
protective suits. Furthermore, especially in missile loading, arming, and validation, intricate close
tolerance dexterous maneuvers are required. These procedures are exceedingly difficult with multi¬
layer gloves.
In 1988, Battelle Memorial Institute in Columbus, Ohio, completed an exhaustive report on
the possibility of using robotic manipulators for performing some or all of the ICT tasks [Smith].
The report recommended near- and far-term scenarios for using robotics to reduce crew exposure to
the CB agents, and to expedite the automated process by concurrent execution of the tasks involved.
Of particular importance to this work, the report recommended an autonomous boom to
refuel the aircraft, taking advantage of the increased flow rate of the aircraft’s aerial refueling port
over the conventional ground refueling port. Refueling would otherwise follow Air Force guidelines
for “hot-pit” refueling (engines running). Automated ordnance loading would be facilitated by
smart robotic “jammers;” vehicles which could be operated by remote control pendants, and which
could position the missile or bomb in its correct position for mounting and arming. To facilitate the
rest of the ordnance loading procedure, it was recommended that the aircraft and missile be re¬
engineered with “quick-connect” connections, reducing the manual dexterity required.
B. Summer Faculty Research Program
In the past, my own work has focused on the interpretation of noisy sensor data in the
intelligent perception of complex objects by robotic manipulators. I am also in the process of
98-3
investigating the properties and availability of various sensors in order to outfit my own robotics
laboratory at Virginia Polytechnic Institute and State University. It is therefore appropriate that I
assist the Flight Dynamics Laboratory in the sensor selection, signal processing, and interpretation
of sensor data in the context of robotic solutions to the aircraft turnaround problem. The particular
tasks are described below.
II. OBJECTIVES
Assuming that the near-term robotic turnaround functions recommended by the Battelle
report are to be implemented, it is the purpose of this effort to examine technical issues in the
feasibility of robotic implementation. In particular;
1) Given the current technological state of robotics techniques and the available commercial
robotics and automation hardware, recommendations will be made as to the practicality of
some of the planned robotic turnaround functions,
2) In the laboratory, a concept demonstration cell has been set up for illustration of the
fundamental aspects of robotic turnaround. Based on tht evaluation of the practicality of the
final implementation, long and short range equipment acquisitions will be analyzed and
recommended.
3) With sp)ecial attention to the particular difficulties inherent in noisy, outdoor/hangar
environments, sensor selection and signal processing techniques will be suggested.
4) For the sensor-intensive application rnvisioned, it is likely that muhi-sensor fusion techniques
will be required for practical implementation. These techniques will be examined from the
current literature, and suggested methods will be chosen, given the sensor assortment
previously selected.
5) In conjunction with the multi-sensor pick-up and processing problem above, issues in robotic
perception of geometry and avtomated e.xploration will be examined for possible use in the
robotic aircraft turnaround fuactions.
The remainder of this report will present a comparison of possibly applicable sensor
technologies. Special attention is paid to features relevant to the rapid aircraft turnaround problem,
both in the lab and ia the field. Die to space limitations, detailed analysis of processing techniques
and some other techiucal data is givei as reference to available literature only.
984
III. SENSOR TECHNOLOGY
A. Ultrasonic Sensors
In situations where lighting conditions are unfavorable for visual inspection, some other
sensing modality is necessary. One popular type of non-contact sensor for such uses is the
ultrasonic ranger. Ultrasonic ranging systems can give very accurate measurements with relatively
compact and inexpensive equipment. Drawbacks include poor directionality and iioise immunity.
Several companies market ultrasonic sensing equipment, but the basic hardware
configuration is similar for all. In fact, many diHerent sensing packages are based on the same
Polaroid electrostatic transducer. The Polariod transducer is light and portable, consisting of a thin
(gold) metallic sheet mounted in a disc-shaped plastic housing, .328 inches thick and 1,69 inches in
diameter. The necessary driving and signal processing electronics as well as the portable battery
packs are separate and attached by wires.
The sensor operates by emitting a “chirp,” lasting approx, 1 microsecond (ms), containing
56 pulses at 300 volts. The pulses are at frequencies ranging from 49.1 to 60 kHz. (This is to avoid
possible cancellation of certain frequencies by surfaces with special topographical characteristics.)
The transducer then acts as a microphone, picking up the echo from its own chirp. The elapsed
time between emission and detection is used with the speed of sound to compute the length of the
echo’s path. Nominal operating conditions in the Polaroid specification sheet are: temperature, 32 -
HOT and relative humidity, 5%-95%. The Polaroid ultrasonic sensor experimenter’s kit includes
circuitry which enables the unit to measure distances in the range 0.9 to 35 feet.
Ultrasonic ranging systems are relatively simple and inexpensive ($200 • $1000) to
implement. The obvious application for robotic environments is in object detection for automatic
collision avoidance. This is particularly true for mobile robots and autonomous guided vehicles
(AGV’s), in which case the sensors operate as a “seeing eye dog” for the robot. Difficulties arise in
the implementation on articulated robots due to the presence of the arms themselves. Because of
the wide field of view of the sensors, parts of one arm might be identified as an obstacle by a sensor
mounted on another arm. Furthermore, if multiple sensors are used, the signals transmitted from
one sensor might be erroneously received and interpretted by a different sensor.
In the rapid aircraft turnaround operations, such detectors might be employed to detect
obstacles, both static and dynamic. For example, a static obstacle is presented by the wing itself
when the required operation is fuel vent verification, since the current scenario has an overhead
robot arm reaching under the wing to check vent flow. Dynamic obstacles are presented by moving
objects, and might include humans, vehicles, and the robot arm itself. An ultrasonic ranger on each
arm segment could detect imminent collision and activate obstacle avoidance algorithms.
98-5
1. IMPLEMENTATION CONSIDERATIONS - As shown in [Polaroid], the beam shape of the
t-
sonic, chirp has a primary lobe approximately 20' wide. Detection circuitry can be readily designed
so that objects of width only .007^^ [Contaq] can be sensed. Therefore, small objects can be detected
in a relatively wide Held, but the objects cannot be further localized. The standard configuration's
therefore most useful for ranging flat objects in a sparse field. Cluttered environments are not easily
mapped, since the signal received indicates only the distance to the nearest object. This situation
can be improved upon with the use of ultrasonic horns, which effectively narrow the beam width to
improve directionality (see datasheets from Massa Products Corp., Hingham, MA 02043). These
structures are available commercially and effectively narrow the beam to less than 10'. A similar
effect can be produced with the use of higher pulse frequencies, which disperse less than the
approximately 50 kHz pulses normally used. Some improved imaging can also be achieved by using
two separate transducers or with scanning sonar and extensive signal processing and filtering
[Beckerman], but this process might be too slow for real-time robotics operations.
The operating temperature range mentioned above is to assure nominal accuracy of
approximately 1% for the detection circuitry. In reality, the useful operating range is somewhat
larger, with some degradation in accuracy. Application notes from the Contaq Technologies
Corporation (Bristol, VT) also mention the possibility of error due to object temperature. The
problem is actually a result of air turbulence due to thermal currents near the surface of the object.
A surface which is correctly ranged when cold might go undetected when hot, since the reflected
beam is dispersed and refracted by the turbulent air. Of course, in a fiight-line situation, there may
be considerable air currents due to nearby jet engines, and these currents may affect the accuracy of
the reading.
Another obvious possible source of error in the fiight-line environment is the jet engine noise
itself. While the noise developed from the jet engine depends on a great many factors, including
nozzle diameter, ambient temperature, nozzle air flow velocity, and angle from the jet axis, the
sound power levels can easily e.xceed 140 dB near a turbojet. Peak power occurs in the 150-1200 Hz
band, with considerable roll off past 10 kllz [Beranekj. Most commercial noise meters have no
appreciable frequency response above 10 kHz, especially so at 50 kHz, where the ultrasonic pulses
are. In the ideal case, the ultrasonic pulse and the jet noise are sufficiently separated in frequency
that interference would be unexpected, or at least easily filtered. However, the power level of the
hot-pit servicing environment is so high that the physical characteristics of the transducer itself may
he altered, changing the frequency response even at high frequencies through a nonlinear effect. As
mentioned, the transducer is composed of a thin gold film, and is not very robust to the mechanical
shock of jet blast.
98-6
The sonic reflectance properties of a surface may also affect sensor readings. Flat, smooth
surfaces should be oriented perpendicular to the beam path for best readings. The Conteiq notes
suggest 10° or less deviation from this perpendicular, but note that sufficiently rough surfaces should
reflect enough of the beam in a wide path that detection is possible and ranging is accurate.
2. EXPERIMENT - On Friday, June 30, 1990, the polaroid ultrasonic experimenter’s kit was tested
on a USAF C-130. The flight-line visit was ostensibly to observe cargo loading procedures, and was
not an ICT environment, nor were there jet engines in operation in the vicinity. However, the
auxiliary power unit (APU) on the aircraft was running and produced high sound levels and hot
exhaust. The purpose of the ad hoc experiment was to test the accuracy of the sensor in the vicinity
of the power unit. The unit was held by hand and pointed directly at the skin of the aircraft,
behind the APU on the left side of the fuselage.
In the range of one to five feet, the sensor performed well, giving stable readings of the
distance to the skin of the aircraft. Between six and ten feet, the sensor gave unpredictable results,
showing stability for brief periods only (less than 10 seconds). It is in. this range that the exhaust
from the APU began blowing eratically into the path of the sound beam. After 10 feet, the sensor
did not give stable readings of the range, despite the 35 ft. capability of the sensor and circuilry. In
this range, the sensor’s beam path was persistently interrupted by the hot exaust. With the APU’s
(and hence, the exhaust) off, the sensor had no discernible difficulty reding ranges of greater than
six feet.
B. Laser Range Finders
For remote sensing purposes similar to those for which the ultrtsonic range sensors were
considered (obstacle detection, etc.), optical range sensors are also available. Most of the applicable
optical range sensors use LED lasers (often near infrared: invisible to the human eye) and
triangulation techniques to determine the distance to an object. Like ultrasonic rangers, these
devices are compact (commonly smaller than a pack of cigarettes, sometimes as small as a
matchbox), and have e.\cellent resolution (10 - 50 /rm).
Unlike ultrasonic sensors, though, they are highly directional. They operate by shining a
laser spot on an object and measuring the location of the spot vith a light-sensitive array. They
require mounting at a nominal distance from the object, called a “stand-off,” ringing from 1 in. to
over 3 ft. Around this distance, displacements of ± 1 to 30 in. can be measuied (approximately,
depending on sensor model and configuration/electronicsi. The primary advantaies of these sensors
are i) they are immune to noise, air currents, and temperature variations; ii) multiple sensors laser
range finders do not interfere with each other’s operations, as ultrasonic sensors will; iii) they are
98-7
comparable in price to ultrasonic sensors ($300 - S600); and iv) they are resonably robust
mechanically, being sealed packages which can be made resistant to CB agents. The disadvantages
are: i) they require opaque surfaces which are not e.xcessively shiny; ii) they require approximately
the same ambient temperature and humiuity conditions as the ultrasonic sensors, which precludes
rainy days: and iii) they are technologically more sophisticated and complex than ultrasonic
sensors, although they are still self-contained and are provided with the interface electronics.
C. Microwaves
A new technology which combines the desireable features of both the ultrasonic sensors and
the optical rangers is the microwave ranger. These devices operate on the same basic principal as
the ultrasonic sensors, but use high frequency (>1 GHz) electromagnetic waves rather than sonic
waves. This technique provides better resolution and immunity to environmental factors than the
ultrasonic devices, and have greater range than optical sensors. A factor to be investigated in the
use of microwaves, though, is possible interference with aircraft and/or flightline electronics and
communications. The devices are commercially available (General Microwave Corporation,
Amityville. NY) and warrant further investigation.
D. Vision and Infrared Imaging
Conventional imaging systems using vidicon technology transform visible light reflected
from a three-dimensiona. scene to the two-dimensional plane image. The obvious utility of this
sensing modality is the remote locatioi of parts and features within the field of view. A properly
calibrated camera can b« useful in many of the aircraft turnaround tasks.
Primarily, tie imaging system used in the current demonstration facility and in the near-
term automated scenario uses visual images for two purposes: to inspect fittings and surfaces for
defects, wear, or damage; and to fix positions of workpieces such as the aerial tefueling port, the
grounding receptacles, the nose gear, the robot end effector itself, and visible obstacles. With
overhead moimted tameras and vision processing software, the incoming aircraft can be identified
and inspected before the refueling ports and other necessary workpoints are identified in three-space.
In tie demonstration facility, a charge-coupled device (CCD) camera is also mounted to the
last link of the robot, so that the local scene is always accessible. While vidicons are electron tubes,
fundan^ntally cathode ray screens operated backwards, CCD’s are solid-state devices fabricated on a
sem^’cnductor chip. This facilitates miniaturization and remote operation. Furthermore. CCD
ca-ieras have high resolution and low power consumption, making them ideal for robotic
pplications. where they have gained wide acceptance.
98-8
Whereas visible light occupies the wavelengths 0.4 - 0.75 fiva in the electromagnetic
spectrum (approximately the same band available from vidicon camerM), CCD cameras have useful
sensitivity from 0.4 - 1.1 fixa. These higher wavelengths extend into the “near” infrared region,
suggesting that electromagnetic emissions from hot objects might be visible with CCD cuneras.
Indeed, CCD cameras are useful in detecting hot objects whose temperatures are > 600*C. There
are, however, many tasks in the turnittouh^l operation for which temperature sensing might be
Useful, especially if the range of ietflpeiatUre resolution were extended to include objects cooler than
60(rC. Such sensing systems ate available in the form of thermal imagers.
Thermal imaging systems ate common in the range 2.0 • 6.0 /im. These wavelengths
correspond, roughly, to the range of wavelengths emitted by common objects (not necessarily ideal
black-bodies) with temperatures of -20* to 200(KC (-4* to 3632*F). A typical system from Hughes
Electronics Company, the Systems Series 2000, operates at 2.0 - 5.6 ^m, can focus in the range 10
in. to infinity, has a 15* x 10* field of view, and plots temperature in a 16 color coding scheme. It
can resolve temperatures within the range -20° to 2000°C within ± 0.5°C. Feature and options allow
frame grabbing, isotherm viewing and “thermal filtering” (selecting only objects of particular
temperatures for viewing), and magnetic storage on videotapes or hard disk recorders. It consists of
an indium-antimonide detector housed in an argon gas coolant.
The possible uses of such a device are numerous:
• Initial inspection of the aircraft includes a check for brake temperature. If the brakes can
melt a crayon designed to melt at 37rC (700T), hot pit refueling is unsafe because of the risk
of spilled fuel ignition. If they are over 649°C (1200T), they are at risk of igniting the
hydraulic fluid, and present imminent danger of brake fire. A thermal imager could detect
brake temperature before any turnaround operations are performed.
• Throughout the refueling process, constant surveillance must be maintained on the fuel tank
vent to ensure that air is properly venting, indicating proper fuel delivery rate. )?or fuel
which has been stored in underground tanks, there is likely to be a temperature gradient
between the venting vapors and the surrounding air. A gradient would also result from
evaporation of the vapors, which will also be the case for liquid fuel spills onto the tarmac.
Such a gradient would surely be within the 0.5° resolving capability of the sensors.
• A well established procedure among chemists is the detection of chemical element through
their electromagnetic emission and absorption properties. It was reported in [Hudson] that
this technique has been used for the detection of fuel leaks as well as the presence of poison
98-9
gases. For e.xample, it was found that hydrogen gas absorbs reidiation at a wavelength of 3.4
fim. Therefore, if spectral analysis of the scene is suspiciously devoid of this wavelength, a
hydrocarbon fuel leak might be suspected. The Hughes system demonstrated in the
laboratory on July 27, 1990 did not have this capability.
• Aside from detecting materials with temperature in the extreme ranges, the imaging system
can also “see” objects with “normal” temperatures, so that some of the more conventional
visual imaging tasks might be accomplished with a single imaging system; for e.xample,
coordinate fixation of the aircraft and localization of the refueling port. As stated in (Lloyd),
“Thermal imagers are superior in performance to other types of passive-sensing electro-optical
imaging devices when operability at any time of the day or night and under all weather
conditions is the primary consideration.” A problem always associated with visible
wavelength imaging in military environments is the necessity to detect camouflaged objects.
Since it has already been determined in the laboratory that a camouflaged aircraft is difficult
to detect on a matching runway, and since a metallic plane returning from a recent sortie is
certain to have a different temperature than the service area floor or ground surface, thermal
imaging might be a valuable option.
Although the thermal imaging systems may have considerable advantages over the CCD or
vidicon cameras, they also have their drawbacks. Because the sensors are temperature-sensitive,
they must be cooled well below the temperature of the objects they image. This need is analogous
to the difficulty in observing visual images with a bright light shining in one’s eyes. Most thermal
imagers are cooled with cryogenic fluids such as argon, dispensed from cylinders which last no more
than several hours (system cost; > S25,000). There are, however, thermoelectrically cooled
systems which require only continuous electrical power, although at a higher cost ( -- !?50,000). The
cooling systems also add some bulk and fragility, eliminating the possibility of mounting on small,
fast robot arms. Gantry mounting or mounting on large, slow moving arms is a feasible option.
In addition, since the wavelengths received are considerably longer than visible light, spatial
resolution is diminished, especially at the low temperature range. The most inexpensive systems are
limited to a small number of colors ( ~ 16) in the display, so representation of large temperature
ranges in one image will cost considerable temperature resolution. The Hughes systems can,
however, change the display range and bias on-line, so selective smaller temperature gradients can be
presented. Hughes also offers 128 and 256 color models, again at a higher price. Complete imaging
systems are available, which include the image processing techniques common to many visual
imagers (image subtraction, filtering, etc.).
98-10
Finally, it should be noted that thermal imagers are not cameras. \t are thermal, sensors
whose inputs are scanned in from the field of view by mirrors (Lloyd). They are considerably slower
than CCDs, having time constants on the order of 100 ms.
E. Force and Tactile Sensors
In the event that desired visual information is unavailable due to poor lighting, occlusion, or
surface refiectance problems, surface features of any object (including the refueling hose and aircraft
surface), touch sensors can be used instead. The possible uses for force and other touch sensors are
many, and include: grip verification and force control for the end effector, position sensing through
detection of wheel contact on the floor or ground, surface e.'cploration and feature detection on the
plane’s surface, perhaps for increased aircraft identification confidence or location of visually
occluded connectors and ports.
A difficulty in the implementation of tactile sensors is commercial availability. Most
research in this area is in development of the sensors themselves, so that selection of a commercial
device is difficult, especially tactile array sensors. Luckily, in the large scale robotics and
manipulation issues necessary here, sensitive sensor arrays are not crucial. E.'camination of
commercial products reveals three types of sensors which are ine.xpensive and useful: the force
sensing resistor, the force/torque sensing wrist module (a strain gage device), and the remote center
compliance with lockout.
The force sensing resistor (FSR) is available from Interlink Electronics (Santa Barbara, CA).
It is a film of semiconducting material whose resistance decrease', as a power of the applied force.
The manufacturer’s information bulletins suggest that while the force sensing capability is useful,
accurate force sensing should be left to strain gage-type devices. The FSR is thin, fie.xible,
mechanically robust, and versatile. It is available in an “XVZ” configuration, which acts as a
contact location sensor with resolution of .002// in the X-Y plane, as well as a force sensor along the
Z a.xis. The basic sensor can be used on gripper faces to monitor gripping, contact force and object
weight. As an XYZ sensor, it becomes useful for object e.xploration and recognition as discussed to
.some extent in [Bay, Bay & Ilemaniij. In that work, it is stated that the signal processing
techniques developed are most useful when accurate position sensing is available, but when force
sensing is relatively inaccurate. This is exactly the situation with the Interlink sensor. Detailed
data and customization information is available from Interlink, and a starter designer's kit will
provide an assortment of 18 sensors and application notes for S79.95.
EXPERIMENT - Two sample FSRs were obtained from Interlink, and the mass-resistance profile
shown in Figure 1 was obtained. It should be noted that the curve is a log-log plot, demonstrating
98-11
that the sensor is a liighly nonlinear device. A further problem discovered in laboratory
experimentation with the sample FSRs was “creep.” Though not specified in the manufacturer’s
data sheet, it was found that over a period of appro.\imately ten minutes, the resistance of the sensor
drops with a constant mass applied. It does decay to a constant and repeatable value, but appears
noisy and unpredictable during the transients. At the time of this writing, e.xperiments were in
progress to reject the creep, either through electronic processing of the measurements, mechanical
damping, or software compensation.
The force/torque sensing wrist senses the resultant forces and torques on the end effector or
hand. It consists of as many as 16 strain gages mounted on each side of metallic cross-bridges
which attach the hand to the wrist. With this redundancy, simple signal processing techniques can
easily resolve a .3-axis torque and a 3-axis force (see Figure 2). The devices are widely used in
robotics, since Joint torques are difficult to measure and are invariably contaminated with nonlinear
disturbances such as gear flexibility, backlash, hysteresis, and friction. In the aircraft turnaround
operation, such a wrist will be necessary not only to monitor workpiece forces, but also to monitor
contact forces, since forcing a part contact with the aircraft skin should be tightly monitored and
controlled. Most force/torque sensors are provided with a dedicated controller, which provides the
somewhat complex operation of strain gage signal pickup, amplification, conditioning, and response
to requests for force/torque information from the host computer.
The remote center compliance (RCC) is a well known device intended to facilitate
manipulation tasks, especially small parts mating. In principle, it provides a spring-dashpot
combination along all six euxes of hand motion, and is constructed so that it “gives” in reaction to
applied force or moments. Inaccurate parts mating, for example, the peg-in-the-hole task, is
accomplished through passive compliance rather than active control. This function will be useful for
any pin insertion or latching tasks, including aircraft electrical grounding and refueling nozzle
insertion. During refueling with a robotic boom or gantry robot, the function of the RCC is the
same as that of the flexibility in the hose during an actual aerial refueling.
The “lockout” feature simply locks the compliance to reduce wobble and undesired flex
during unobstructed and non-contact moves, thereby also facilitating use of a force/torque sensor in
roniunction with an RCC. RF’C’s, as well as force-sensing wrists, are available from Assurance
Technologies (Garner, .N'C) or PF.\, Inc. (New Berlin, WI). Force/torque sensors cost S4000 -
S10,000. depending on the payload capability ( ~ 15 Ib./lO in.-lb. ~ .350 lb./2100 in.-lb.), and
RCC’s are about S500. or SlSOO with lockout. Some robotics literature cribes a combination of
the two devices in one module, the “instrumented RCC.” but a commerciallj available product was
not found.
;
98-12
IV. RECOMMENDATIONS
The following paragraphs summarize the final recomtnendations of the suriimer
investigation, particularly with regard to the objectives listed above.
A. SENSORS - Because of the problems associated with ultrasonic transmitters/receivers as
mentioned in section III. A. they are probably not suitable for multiple mountings on an articulated
arm. Their use should be limited to range estimation by single mobile robots. Because of stringent
safety requirements dictating accurate range sensing and obstacle avoidance, laser range finders are
better suited to the flightline environment. Microwave devices should be further investigated.
Because of the low tolerances allowed, it is strongly recommended that the laboratory upgrade to
industrial grade robot arms, and equip them with force/torque sensing wrists and RCC devices.
Contact applications around critical <and expensive flightline equipment justify their cost. Tactile
and gripper-mounted array sensors beyond basic force sensing should be implemented in the
laboratory only if artificially intelligent autonomous manipulation or shape description (see
D. ROBOTIC PERCEPTION . . . , below) is desiredi In this case, preceding techniques are widely
available in the voluminous robotics and image processing sources literature references. The CCD
overhead and gripper-mounted vision systems are adequate for laboratory study and should be useful
on the flightline. Their use should be continued, and augmented with additional vision processing
software. An actual flightline operation might also benefit from thermal imaging, although the cost
of these systems is prohibitive in the laboratory. A useful comparison of sensor technologies was
found in [Espiau] and is summarized in Figure 3.
B. AUTOMATION of SELECTED TURNAROUND FUNCTIONS - As recommended by the
Battelle report, refueling is the primary turnaround function currently suited to full automation.
Bomb and missile loading is currently too intricate a task to be fully automated, although gross
ordnance transfer operations are feasible with semi-autonomous guided vehicles. Automated
reloading will not be a candidate for full automation until “quick connect” fittings are implemented,
implying substantial re-design of the pylon, retaining, and arming structures on the plane and the
missile/bomb. Automated visual/thermal inspection is a technologically feasible application for
mobile robots and should be considered for integration with the refueling apparatus. Wheel
chocking is best performed with re-design of the chocking device, perhaps with inflatable bladders
over which the aircraft might be taxied prior to inflation. Grounding, umbilical connection, and
initial inspection can be accomplished with the same overhead apparatus as the refueling operation,
but might imply identifiable visual markers on the surface of the aircraft. Ground plug insertion
and fuel hose mating requires some passive compliance device such as an RCC.
98-13
■C. SIGNAL PROCESSING - It wm diKovered from the literature that current multi-sensor- fusion
techniques are driven uniquely by the particular asirortment of sensors in the application. .No
accepted, unified multi-sensor integration paradigm e.^ists, and should be the topic of further
research. Sensor subsystems should become modular (“plug-in and go'); Until the time that this
becomes possible, sensor fusion software will be custom-written for ea.ch application. Multi-sensor
integration for problems similar to the refueling task are considered in the literature [Abidi] and
would make a good guide for refueling operations. Obstacle avoidance procedures are common in
robotics literature and will be necessary here. Especially relevant and potentially problematic is the
problem of avoiding collision between the refueling; boom and the aircraft. This is inherently a
three-dimensional mapping, and avoidance problem, as opposed to the much easier planar problem.
•A promising approach is to create a. ficticious repulsive force exerted by the surface of any obstacle
detected by the range sensors [Cheung), although this is most effective for fixed obstacles. The
obstacles present then create an artificial “force field,” repelling each link of the robot arm away
from object surfaces. This is an elegant and passive technique and requires no global planning and
mapping.
D. ROBOTIC PERCEPTION AND SHAPE DESCRIPTION - In remote locations or bare base
operations, it has already been determined that camouflaged aircraft may be difilcult to localize by
camera. This may be problematic if insufficient data is available to precisely fix the physical
coordinates of the various aircraft fixtures. To remedy this situation, salient features of the aircraft
should be accurately fi.xed with .other sensor inputs. First, if the inflatable bladder or floor-mounted
automatic chock is used, this device could also be instrumented to provide approximate resting
location of the aircraft tires. Then, if the end effector is equipped with either range or touch sensing,
the surface of the aircraft can be explored, and measured geometric features matched with the
known characteristics of the aircraft model. Sensor information useful for this procedure includes
range, surface normal direction, curvature, and even force (for constrained motion on unknown
surfaces). This information can then be fused with visual information so that a surface model can
be fitted to the data, and coordinate references can be estimated. The procedures specified in (Bay
& Ilemamij will accomplish this estimation and the exploration tasks are being investigated (Bay).
Other promising approaches include neural network classifiers (Norwood). It is recommended that
hardware experimentation with tactile shape estimation be conducted on aircraft models in order to
verify the theory in the laboratory.
E. OVERALL - The authors of the Battelle report (Smith) used a technique of assigning scores to
various turnaround functions based on feasibility, potential time savings, etc. The combined score
then ranked the functions in order of strength as candidates for full automation. Curiously, the
98-14
functions of bomb lo^^ing and AIM-9 missile loading ranked higher thaii refueling, but refueling was
nevertheless chosen as the best candidate for immediate prototype development.
Based on the analysis of sensor equipment and manipulator hardware available, refueling
does appear to be the best candidate for automation (along with the associated necessities of
grounding, chocking, etc.). Next, noncontact inspection should be implemented. This can be done
with currently available remote sensors. The primary difficulty with inspection is the comparative
time required to scan the entire surface of the aircraft at high resolution, something humans, do well.
The primary advantage is the ability of the robotic system to exploit sensors (X-ray, etc), which can
detect minute or invisible flaws.
Autonomous bomb and missile loading should be left for further study in the future. While
mounting and arming the weapons requires intricate maneuvers and fine dexterity, handling the
weapons requires high strength and close supervision. This wide range of capabilities is difficult to
build into an autonomous system, and it requires technology currently unavailable. Although non-
autonomous force amplifiers and human-in-the-loop mechanical assistance concepts [Kazerooni] show
promise, these operations are so safety critical that more development and testing time is required.
It has been said that robotics can be defined as the marriage of artificial intelligence and
movable physical objects. If this is true, the current state of industrial applied robotics is better
defined as automation, which implies little intelligence. Although crew e.xposure to CB agents and
physical workload advantages are immediately realizable, time and cost savings- currently are only
appreciable in repetetive tasks requiring little sensory perception and rational decision-making.
Since the comprehensive rapid aircraft turnaround is such a complex, sensory intensive, low
tolerance problem with little room for error, considerable further research is necessary.
98-15
I would like to express my appreciation to the staff of the Flight Dynamics Laboratory^ Vehicle
Subsystems Division, WRDC/FIVMB. In particular, I would like to thank Dr; Mangal Chuwla, Mr.
David Perez, and Dr. Arnold .Mayer for the opportunity to participate in the summer faculty
research program. I am also indebted to the rest of the robotics staff at FIVMB for their assistance
in -the research contained herein, and for the wealth of background information and logistics support
necessary for the work.
Universal Energy Systems, Air Force Systems Command, and Air Force Office of Scientific Research
also deserve much credit for initiating and administering university-government cooperation and
collaboration through the Summer Faculty Research Program.
98-16
REFERENCES
Abidi, M. A., and R. C. Gonzalez, “The Use of Muitisensor Data for Robotic Applications,”
IEEE Transactions on Robotics and Automation, vol. 6, no. 2, April 1990, pp. 159-177.
Bay, John S., “A Fully Autonomous Active Sensor-Based Exploration Concept for Shape
Sensing Robots,” subm. to IEEE Transactions on Systems. Man, and Cybernetics. May 1990.
Bay, John S., and H. Hemami, “Dynamics of a Learning Controller for Surface Tracking Robots
on Unknown Surfaces,” Proceedings of tfis 1990 IEEE International Conference on Robotics and
Automation. Cincinatti, OH, 1990, pp. 1910-1915.
Beckerman, M., and Oblow, E. M., “Treatment of Systematic Errors in the Processing of Wide-
Angle Sonar Sensing Data for Robotic Navigation,” IEEE Transactions ©a Robotics and
Automation, vol, 6, no. 2, April 1990, pp. 136-145.
Beranek, Leo L., ed.. Noise Reduction. New York, McGraw-Hill, 1960.
Cheung, E., and V. Lumelsky, “Motion Planning for a Whole-Sensitive Robot Arm
Manipulator,” Proceedings of t]T£ IEEE International Conference oa Robotics and Automation.
Cincinatti, OH, 1990, pp. 344-349.
Contaq Technologies Corporation, 15 Main Street, Bristol, Vermont, 05443, (803) 453-3332.
Espiau, B., “An Overview of Local Environment Sensing in Robotics Applications,”, in Dario,
Paolo, ed.. Sensors and Sensory Systems for Advanced Robots. .New York, Springer-Verlag,
NATO ASI series vol. 43, 1988.
Hirzinger, G., and J. Dietrich, “MultisensoryFeedback Including Cooperative Robots,” in Dario,
Paolo, ed.. Sensors and Sensory Systems for Advanced Robots. New York. Springer-Verlag,
NATO ASI series vol. 43, 1988.
Hudson, Richard Jr., Infrared Svstem Engineering. New York, Wiley-Interscience, 1969.
98-17
Kazerooni, H., “Human- Robot Interaction via the Transfer of Power and Information Signals,”
IEEE Transactions on Svsterns. Man, and Cybernetics, vol. 20, no. 2, March/April 1990, pp.
450-463.
Lippmann, Richard P., “An Introduction to Computing with Neural Nets,” IEEE Acoustics.
Sheech. and Signal Processing Magazine. April 1987, pp. 4-22.
Lloyd, J. M., Thermal Imaging Systems. New York, Plenum Press, 1975.
Norwood, John, Peter WeilandJ and J. B. Cheatham, “Robotic Navigation from Local Sensory
Data Using Neural Networks,” Proceedings of tl]£ International Robots ^ Vision Automation
Conference. Detroit, Michigan, 1990;. pp. 7-130 — 7-138.
Polaroid Corporation, Ultrasonic Ranging System. Manual for Ultrasonic Ranging System
E-xperimentation Kit.
Casals, Alicia, ed.. Sensor Devices afli^ Systems fo£ Robotics. New York, Springer-Verlag, NATO
ASI series, vol. 52, 1988,
Smith, John L., et al.. Study of Robotic Concepts for Aircraft Turnaround. Technical report no.
AFWAL-TR-88-3058, Battelle Memorial Institute for Flight Dynamics Laboratory, AF Wright
Aeronautical Laboratories, WPAFB, OH, 1988.
98-18
Resistance vs. Mass for Interlink FSR Samples
10^2 - -
“ 7/12/90
;(y> - ^ - uj — ^ - -
10' 10- '0-' W
Mass Applied (grams)
Figure 1. Resistance vs. mass applied for the sample Interlink FSR. Note the log-log scale and
the outlier, probably due to the creep discussed in section E.
/
Figure 2. Structure of the strain-gage force/torque <onsor representative of commercially
available products. External processing e.xlracis directional iiiformation from the bridge-
connected strain-gage pairs [Hirziiigerj.
Copy available to DTIC does not
’permit fully legibk* d'j-'tjon
98-19
98-20
Copy available to DTIC does not
permit fullv Koibk* c p d>i''t'.on
1990 USAF-UES SUMMER FACULTY RESEARCH PROGRAM
Sponsored by the
AIR FORCE OFFICE OF SCIENTIFIC RESEARCH
Conducted by the
UNIVERSAL ENERGY SYSTEM, INC.
Prepared by:
Academic Rank:
Dr. Franklin E. Eastep and Anne Stephenson
Professor
Graduate Student
Department and
University:
Research Location:
Aerospace Engineering
University of Dayton
Flight Dynamics/ Laboratory, Structures and
Dynamics
Division, Analysis/Optimization Branch
USAF Researcher: Dr, Vipperla B. Venkayya
Date:
1 November 1990
contract No . :
F49620-88-C-0053
INFLUENCE OF STATIC AND DYNAMIC AEROELASTIC
CONSTRAINTS OF THE OPTIMAL STRUCTURAL
. DESIGN OF FLIGHT VEHICLE STRUCTURES
Franklin E. Eastep, Ph.D.
Department of Aerospace Engineering
University of Dayton, Dayton, OH 45469-0227
ABSTRACT
This investigation focused upon' the structural weight
optimized design of a fighter-type wing of low aspect ratio using
ASTROS. The optimal weight redesign of a preliminary finite
element model representing the wing structure is obtained with the
constraints on strength, control reversal and flutter imposed using
both subsonic and supersonic aerodynamic theories. It is
demonstrated that the optimization capabilities of the ASTRO
procedure are well suited for the preliminary structure design
environment. ASTROS gives to the structural designer the
capability to develop unique solutions to the design problem facing
flight vehicle structures with the many constraints.
Recommendations are made to Include a transonic aerodynamic
formulation with ASTROS for the structural design of a flight
vehicle over the entire Mach number regime.
ACKNOWLEDGEMENT
The author would like to thank the Air Force Office of
Scientific Research and Universal Energy Systems for providing him
with the opportunity to spend a worthwhile summer of 1990 at the
Air Force Flight Dynamics Laboratory, Wright Patterson AFB, Ohio.
He would like to acknowledge the laboratory; and in particular, the
Analysis and Optimization Branch, for the opportunity for the
exchange of ideas with others working in the area of weight
optimization of flight vehicle structures.
Finally, he would like to thank Dr. Vipperla B. Venkayya for
suggesting this area of research and for his collaboration and
guidance. Additionally,, he would like to acknowledge many helpful
discussions with Mark French, Ray Kolonay and Dr. Fred Striz.
99-2
I . INTRODUCTION
An aircraft structural designer must consider aeroelastic
instabilities (i.e. flutter, divergence and control reversal) in a
addition to the strength requirements for the structural design of
high performance aircraft. In particular, he must design a
structure such that the maximum operational flight velocity is less
than (by 15%) the velocity of the critical aeroelastic instability
while still insuring satisfactory strength at the velocity of the
critical aeroelastic instability. The critical aeroelastic
instability is defined to be the least value of the flutter,
divergence or control reversal velocities. In addition, the
structural designer desires to adjust the structural sizes to
minimize the structural weight while insuring that the maximum
flight velocity is less than the critical aeroelastic instability
velocity.
In recent years, structural optimization as needed and used by
the aerospace industry has expanded in scope to include such
additional disciplines as static and dynamic aeroelasticity,
composite materials, aeroelastic tailoring, etc. One of the more
promising multidiciplinary codes presently under development is the
Automated Structural Optimization System (ASTROS) [1-3]. In this
computer code, static, dynamic, and frequency response finite
element structural modules, subsonic and supersonic steady and
unsteady aerodynamic modules, and an optimization module are
combined and allow for either analysis or optimized design of given
aircraft configurations. Interfering surface aerodynamics are
incorporated to handle the aerodynamic modeling of combinations of
99-3
wings, tails, canards, fuselages, and stores. Structures are
represented by fully built-up finite element models, constructed
from rod, membrane, shear, plate, and other elements, static and
dynamic aeroelastic capabilities include trim, lift effectiveness,
aileron effectiveness, gust response, and flutter analysis. The
optimization and aeroelasticity modules of this code were used as
a tool for the structural optimization of fully built-up finite
element wing models in subsonic and supersonic flow with strength
as well as static and dynamic aeroelastic constraints.
First, the performance of the flutter analysis/optimization
module was evaluated against results from the other methods and
codes, i.e., the large scale finite element codes MSC/NASTRAN [4]
and COSMIC-NASTRAN [5], Here, a generic low aspect ratio fighter
type wing model of reasonable complexity (86 nodal points) was
analyzed for flutter by all three codes, then optimized for minimum
weight with strength and various aeroelastic constraints by ASTROS
using a small number of design variables (=5) and design variable
heavy linking. The optimum designs were again analyzed by all
three codes, i.e., free vibration mode shapes, flutter speeds, and
flutter mode shapes were determined and the results from the
different codes were compared. Also, for some of the cases, the
free mode shapes, the flutter speeds, and the flutter mode shapes
were monitored from iteration to iteration during the course of the
optimization, together with the constraint behavior to understand
the performance of the optimization routine especially during steps
in the optimization when various different constraints were active
at the same time. In the investigation, the strength constraints
99-4
were incorporated in the form of a 9-g symmetric pull-up maneuver,
the static aeroelastia constraints were based on control
effectiveness and aileron reversal, and the dynamic aeroelastic
constraint was represented by a minimum flutter speed.
II. OBJECTIVES
The objective of this study is to demonstrate that the
optimization capabilities of ASTROS are well suited for the
preliminary structural design environment. The constraints of
strength^ control reversal and flutter speed are imposed on a
preliminary finite element structural modal of fighter-type wing to
obtain a weight optimized wing structure. It is desired to show
the ability for ASTROS to simultaneously consider many constraints
from each of several disciplines allowing the structural designer
to develop non-intultive solutions of the complex design problem
placed on modern flight vehicle structures.
III. EQUATIONS FOR STRUCTURAL AND AEROELASTIC DESIGN
Consider the swept and tapered thin wing shown in Figure 1
which is typical of a wing investigated herein. The underlying
structure of the wing is represented using finite elements as shown
in Figure 2 which is typical of a wing built-up structure. In
addition, also shown in Figure 1 is an aerodynamic flap located
near the wing tip whose purpose, in conjunction with a flap located
on the opposing wing, is to create a rolling moment for rolling of
the flight vehicle.
Begin the investigation by first determining the static
aeroelastic deformation of the wing produced by the aerodynamic
99-5
forces acting on the wing. The equation of equilibrium for the
bending and twisting of the wing is written in matrix form
conforming to the finite element formulation as:
[Jtliu} = g[A]{u} + g[A*]{a^} + g[A^{p} (1)
where {u} represents the bending and twisting deformation at the
nodal points of the finite element model (a^} is some initial angle
of attack of the wing and { P } is flap setting angle. The
parameter q is the dynamic pressure defined as 1/2 p. Uo^ . In
equation (1) the matrix [K] is the structural stiffness and [A] is
the aerodynamic flexibility matrix representing the aerodynamic
lift and moment generated by the deformation of the wing beyond
that which depends on the initial angle of attack and flap setting.
The equilibrium equation (1) can be written in the usual form for
determining the deformation {u} as
[[k] - g[A]]{u} = g[A^]{a*} + gUntp}. (2)
1. Wing Divergence Velocity
The wing divergence velocity is independent of the initial
wing angle of attack and the flap setting angle. At the divergence
velocity there is a balancing between the aerodynamic force
produced by wing deformation and the structural resorting force
represented by the product of the stiffness and wing deformation.
To determine the divergence dynamic pressure or velocity equation
99-6
(2) is written in homogenous form.
IIK\ - ]{u} = {o}
(3)
Equation (3) is an eigen-problem, from which, the lowest
eigen-value can be determined and represents the divergence dynamic
pressure. Additionally, the accompanying deformation shape, the
eigen-vector can be determined. It should be noted that the
aerodynamic influence metric [A] in general is not a symmetric
matrix which requires special numerical treatment of the eigen-
problem represented by equation (3) because of the possibility of
complex eigenvalues. An inverse-power method was used for the
eigenvalue extraction from equatiron (3), The divergence velocity
can then be written as:
P-
(4)
2. Lift Effectiveness
Lift effectiveness is defined to be the ratio of wing lift
produced on a flexible wing to that produced on a rigid wing at a
defined value of dynamic pressure. The lift effectiveness is
computed at a certain velocity where the flap setting angle, P
is zero. The lift effectiveness is then:
99-7
where the matrix {h} are aerodynamic panel lengths. The wing
deformation vector {u} in equation (5) is computed from equation
(2) with the flap setting { P } equated to zero as:
{u} = g[[lC\ -
(6)
Substituting elation (6) into equation (5) allows one to determine
the lift effectiveness.
=
(7)
>
3. Control-Surface Effectiveness
The purpose of the flap shown in Figure 1 located near the
wing tip is to produce an aerodynamic rolling moment of the wing.
The deformation of the wing surface will produce additional rolling
moments beyond those produced if the wing was rigid. The rolling
moment effectiveness is defined to be the ratio of the rolling
moment resulting from a flap rotation on flexible wing to rhe
rolling moment resulting from a flap rotation on a rigid wing. The
control effectiveness is computed for a specified velocity with
initial wing angle of attack equated to zero in equation (2)
99-8
„C ^xzisid*Mxflex _ ^ {p}-l-C?{A}^[A] id
■' g{A}nAmp}
(8)
where the matrix {h}’*' are products of p^nel lengths and moment arms
to the nodal points of the structural finite element model.
Additionally, it should be noted that the matrix { P } has a non¬
zero value only for elements where the flap is located. Once again
the deformation vector {u} can be determined from the equilibrium
equation (2) but in this case the initial angle of attack is
equated to zero.
id = g[[i^ - gU]]-MA^{p} (9)
Substituting equation (9) into equation (8) the flap effectiveness
for creating a rolling moment on a flexible wing is determined as:
... (h}^[[j] q[a] [[k\ - guj]-^i [A nip)
^ ■ {h}nAn{p} ^ ^
The flap control effectiveness can thus be determined for any
specified value of velocity or dynamic pressure.
4. Control-Surface Reversal
From consideration of equation (10) the possibility exists
that there is a velocity or dynamic pressure where the control
effectiveness is zero. The velocity at which the control
effectiveness is zero is called the control reversal velocity for
the rolling moment produced by a flap deflection on a flexible wing
99-9
is zero and any further increase of velocity will produce a rolling
moment tending to roil the wing in a direction opposite to the roll
produced by a flap rotation on a rigid wing. The reversal velocity
can be determined from a graphical di^p^lay of equation (10) , E' for
various values of dynamic pressure.
5. Flutter Analysis
To determine the flutter speed of a flight vehicle wing shown
in Figures 1 and 2 it is necessary to obtain the equation of motion
representing deformation of the structure when subjected to the
unsteady aerodynamics forces . Two aerodynamic methods are
available in ASTROS, the doublet lattice method for subsonic
velocities and the potential gradient method for supersonic
velocities; unfortunately there is not an aerodynamic method for
transonic velocities. The flutter stability analysis is based on
the p-k method with the equation of motion governing wing
deformation as
pHm +
(o)
(11)
where p is a complex eigenvalue, [M] is the consistent mass matrix
and [Q] = [Q“] + i [Q^] is the complex aerodynamic coefficient
matrix. The flutter speed is determined by plotting the complex
eigenvalue on a v-g diagram.
6 . Design Problem
The concept of optimization involves the alteration of design
99-10
variables (thickness) to achieve desired results. The optimization
goal is to find a vector x of n design variables X;, (i=l,2. . . ,n) ,
which will minimize a multivatiable function f(x) subject to m
constraints:
gr^(x) = Gjix) - Gj o(j = 1,2, . .. ,m) (12)
and side constraints
Xi i i Xi a = 1,2, ,n) (13 )
Here the design variable vector controls the finite element
sizes, the constraints Gj(x) are strength, aeroelastic divergence,
control effectiveness and flutter speed, Gj are the limits, f (x)
represents the structural weight and the superscripts land u
represent the lower and upper bounds.
VI. NUMERICAL RESULTS AND DISCUSSION
Two fighter type wing structural models shown in Figure 3 were
selected to be representative of wing of a low aspect ratio. The
preliminary design model was selected for weight optimization under
strength as well as static and dynamic aeroelastic constraints for
a reasonable sized model to allow parametric investigations. The
large scale model would likely be used in the later phases of the
structural design process.
Initially, the sizes and locations of the structural elements
of a 10 spar, 3 rib wing shown in Figure 3 were selected and as a
99-11
nominal structural model . Additionally) concentrated weights were
placed at the structural nodal points to simulate the non-
structural weights to represent fuel, actuators and store mass.
The aerodynamics modeling for the nominal and optimized structure
was selected to be 36 aerodynamics boxes with 6 chord divisions and
6 span divisions.
1. Nominal Wing Structure
The nominal wing structure was used with ASTROS to determine
the individual stresses in the structural elements resulting from
a 9-g pull-up at a Mach no of 0.85 at sea level. Additionally, the
flutter speed of approximately 32,500 in/sec for the nominal model
was found.
Finally the roll effectiveness for a roll control system was
determined for the nominal wing structure at a dynamic pressure
near the control reversal dynamic pressure. The variation of
control effectiveness and aileron effectiveness of the nominal
structure is displayed in Figure 4. As indicated in Figure 4 the
reversal dynamic pressure for the nominal model was approximately
45#\in2.
As a side investigation to determine the reliability of ASTROS
to predict a valid flutter speed, the flutter speed as predicted
using the p-k method of ASTROS was compared to the flutter speed
predicted using both the k and p-k methods of COSMIC-NASTRAN. As
shown in Figure 5; the v-g diagram of the critical flutter mode of
the nominal structure, all three methods predict exactly the same
flutter velocity even though all three methods yield very different
99-12
damping values at velocities other than at flutter, ih addition;
comparison of flutter velocities predicted by three methods is
displayed in Figure 6 for an optimized structural model (structural
weight less than the nominal model using a preassigned flutter
speed constraint) i The v-g diagram, Figure 6 again demonstrates
the same prediction of flutter velocity by the three techniques and
for the optimal structural model the damping values, g, are
approximately the same for all three techniques at all velocities.
The displayed results of Figures 5 and 6 demonstrate that the p-k
flutter techniques of ASTROS predicts the correct flutter velocity
in the analysis or design phase of ASTROS.
2. Optimized Wing Structure
The fighter wing structural model was resized and optimized
using ASTROS with both single and multiple constraints active at
any given time as shown in Table I. The initial optimal structure
model was obtained for a 9-g symmetric pull-up maneuver with a Von-
Mises stress constraint with prescribed stress yield values. The
following optimization studies were conducted using as minimum
sizes for the individual structural members those obtained from the
above strength optimized model.
Next, the structural model was resized and optimized using a
constraint of an improvement of the reversal dynamic pressure from
45#/in^ to 52#/in^. The increase in the reversal dynamic pressure
was accomplished while the structural weight of the optimized wing
was reduced from 497# (nominal) to 409# (optimal) as shown in Table
I. As a comparison problem with the reversal velocity used as a
99-13
constraint, the structure was resized to yield the same reversal
dynamic pressure as the nominal structure and weight optimized to
reduce the structural weight further to 347#. The corresponding
flutter velocities for these two cases with constraints on reversal
dynamic pressure are indicated in Table I and are 3.23 x 10^ (in/sec)
and 2.97 X 10*(in/8ec) respectively.
Finally the structural model was resized and optimized using
multiple constraints. In this case it was required that the
reversal dynamic pressure be the same as reversal of the nominal
structure and the flutter speed be Increased to 3.1 x 10^ in/sec.
This increase of flutter speed is beyond that obtained when only a
single constraint on reversal pressure was imposed. In this manner
the structural designer has the added advantage of precisely
placing the velocity of certain aeroelastic instabilities relative
to other velocities of aeroelastic instabilities. Here we desired
to make improvements in the flutter velocity which is larger that
the reversal velocity while the structure is weight optimized. In
this particular case the weight of the optimized structure was
obtained as 438#.
The previous optimization studies were accomplished in the
subsonic Mach number regime (i.e. Mg, » 0.85) using a doublet
lattice aerodynamic formulation. Using subsonic aerodynamic
theories resulted in the prediction of aeroelastic instability
velocities in the supersonic regime. To remove this inconsistence
in problem formulation a supersonic aerodynamic formulation, called
a potential gradient method, was used in an optimization study.
The Mach nuiaber was selected to be 1.2 and the structure was
99-14
WEIGHT OPTIMIZED WING WITH VARIOUS CONSTRAINTS
(ALTITUDE - SEA LEVEL)
99-15
resized and optimized with single and multiple constraints also
shown in Table I. The selected constraints for M=1.2 were similar
to the constraints selected for H-0.85 and the weight reductions
for the optimized structure is shown in Table I.
V. CONCLUSIONS
The examples presented in this investigation demonstrate that
the optimization capabilities of the ASTROS procedure are well
suited for the preliminary design environment. Any number of
constraints of strength, divergence, control reversal and flutter
can be imposed on general finite element structural models of
flight vehicles. The ability to simultaneously consider many
constraints from each of several disciplines allows the structural
designer to develop rion-intuitive solutions to the complex design
problem placed on modern flight vehicle structures.
This investigation focused upon the structural weight optimal
design of a fighter-type wing of low-aspect ratio. The finite
element representation of the structure was sufficiently large and
the design variables large enough to serve as a preliminary design
model. The optimal weight redesign of the wing structure was
obtained with constraints on strength, control reversal and flutter
imposed using both subsonic and supersonic aerodynamic theories.
The weight savings for the various imposed constraints is shown in
Table I.
VI . RECOMMENDATIONS
At the present time only the linear aerodynamic theories of
the sub and supersonic Mach number regimes are available for use by
99-16
ASTROS in the optimal weight design of flight vehicle structures.
The critical flight regime for structural design to preclude the
flutter instability is the "transonic" Mach number regime. In this
regime the aerodynamic equations are non-linear and net available
for use by ASTROS in the very regime that is most critical. It is
recommended that a transonic aerodynamic formulation be
incorporated with ASTROS to obtain the weight optimized design of
flight vehicle structures for all Mach number regimes. The
aerodynamic formulation exists at the present time as XTRAN3 being
developed by the AF Flight Dynamics Lab. XTRAN3 is a numerical
solution scheme using finite differencing method to solve the non¬
linear aerodynamic flow equations.
REfBRBMggS
1. Johnson, E.H., and Venkayya, V.B., "Automated Structural
Optimization System (ASTROS), Volume I - Theoretical Manual",
AFWAL-TR-88-3928/I, Air Force Wright Aeronautical, December
1988.
2. Neill, D.J., Johnson, E.H., and Herendeen, D.L., "Automated
Structural Optimization System (ASTROS), Volume II - User's
Manual", AFWAL-TR-88-3028/II, Air Force Wright Aeronautical
Laboratories, December 1988.
3. Johnson, E.H., and Neill, D.J., "Automated Structural
Optimization System (ASTROS) , Volume III - Applications
Manual", AFWAL-TR-88-3018/III, Air Force Wright Aeronautical
Laboratories, December 1988.
4. Rodden, W.P., Editor, MSC/NASTRAN Handbook for Aeroelastic
Analysis, The MacNeal-Schwendler Corporation, Los Angelas;,
California, 1987.
5. "NASTRAN User's Manual", NASA SP-222(08), June 1986.
6. Striz, A.G., and Venkayya, V.B., "Influence of Structural and
Aerodynamic Modeling on Flutter Analysis", Proceedings, 31st
AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics and
Materials Conference, Long Beach, California, April 1990, pp.
110-118.
99-17
I
Hgurt 1. I<l*«lt*t<J Wing PUnform Ceonttry.
Figure 4. Control Aileron Efficiency versus Dynaaic Pressure
(M • 0.85i sealevel).
99-18
> (.Normal
Frecstrcam)
SKINS
(MEMBRANES^
RIBS
(SHEAR
PANELS)
9
10-SPAR FIGHTER WING
MACH 0.85 AT SEALEVEL
— , — , — — , —
-0- NOW ASTROS
NOM NASTRAN-PK ,
/
-
• t
NOk
NASTI
?AN-K
/
t
t
t
t
/
- - -
N— »
<
0 -)
9 ^ 1
y"
‘ V .
_
«
_
o.-l I I I I • ' I I I I I I
0^®
V (IN/S)
Figure 5. Damping versus Velocity of the Critical Flutter
Mode - Nominal Structure.
10-SPAR FIGHTER WING
MACH 0.85 AT SEALEVEL
n
-O- OPT
OPT
— 1 — 1 —
ASTROS
NASTRAN-PK
■
■
■
■
■
ri
■
■
■
■
■
v
1
■
■
■
■
■
■
1
8^'
1
■
■
■
■
■
s
i
■
1
t'*
rt— -
t""'
A — .<
u
f>r
■
■
■
■
■
■
■
0^“° ^oooo
V (IN/S)
Figure 6. Damping versus Velocity of the Critical Flutter
Mode - Optimal Structure.
99-20
1990 USAF-UES SUMMER FACULTY RESEARCH PROGRAM
GRADUATE STUDENT RESEARCH PROGRAM
Sponsored by the
AIR FORCE OFFICE OF SCIENTIFIC RESEARCH
Conducted by the
Universal Energy Systems, Inc.
£IML-SEfflSI
Location of Crack Tips bv Acoustic Emission for
Application To Smart Structures
Prepared by:
Academic Rank:
Department and
University:
Research Location:
USAF Researcher:
Date:
Contract No:
Marvin A. Hams tad, PhD
Professor
Engineering Department
University of Denver
WRDC
Flight Dynamics Laboratory
Wright-Patterson AFB
(Dayton) OH 45433-6553
George P, Sendeckyj , PhD
17 Sept 90
F49620-85-C-0053
Location of Crack Tips by Acoustic Emission
for Application to Smart Structures
by
Marvin A. Hams tad
abstract
The use of commercial acoustic emission (AE) equipment for location
of crack tips in a fatigue test environment was studied. The results from
six channels of commercial AE equipment were compared to those derived from
waveforms obtained by a twp-channel transient recorder, Prior to fatigue
cycle monitoring, the AE wave propagation characteristics were extensively
studied using pencil lead breaks in a center notch. For the fatigue
studies center crack samples of 2024-T351 aluminum were used with hydraulic
grips to eliminate extraneous noise over the whole fatigue cycle. Results
show that waveform approaches are superior to standard AE systems for
location of crack tips. Results also indicated that reopening of a closed
crack generates much more AE than that generated at the crack tips . This
"unclosure" AE may be of potential use to detect cracks and characterize
crack length.
100-2
I . INTRODUCTION
The Fatigue, Fracture and Reliability Group of. the Flight D3mamlcs
Laboratory of the Wright Research and Development Center is interested in
the concept of smart structures. This terminology refers to the ability
to nearly continuously sense conditions and states such as loads and
structural integrity in aircraft structures [Sendecl^j and Paul, 1989].
The inputs of the sensors into properly programmed computers lead directly
to corrective actions thereby by-passing scheduled nondestructive
evaluation. Acoustic emission (AE) technology provides one approach to the
network of sensors and computers necessary for a smart structure.
AE technology -is concerned with the monitoring of stress waves which
are generated by rapid local energy releases in solid materials. The AE
signals are generated by a variety of sources such as impacts, phase
changes, crack growth, friction, inclusion-particle fracture, and other
microdamage sources.
AE technology has a number of useful and potentially unique features
for smart structures •.applications. First, AE is a passive technique in
that stress waves generated at sources throughout a structure propagate to
the sensor location. This feature limits the density of sensors required
to monitor a structure. Second, AE is a whole volume sensing technique.
Third, AE data from multiple sensors can be used t;o locate the spatial
position of the sources within the structure. Fourth, the source mechanism
may be identified by use of advanced statistical concepts with waveform
feature vectors. Fifth, AE is essentially a real time sensing technique.
Sixth, AE is a microscopic technique with high sensitivity. And seventh,
AE is at its very basis a measure of the response of a structure to stress.
II. OBJECTIVES OF THE RESEARCH EFFORT
Since a goal of the smart structures concept is to bypass the usual
scheduled nondestructive inspections, the application of AE would make its
most valuable contribution if crack lengths could be measured by following
the position of crack tips in a fatigue environment. AE technology as
normally practiced in structures has only been concerned with detecting and
100-3
locating the vicinity of a crack. Thus, the smart structures application
requires an extension of the technology to locate crack tips. With this
perspective in mind, the prima^ objective of this research was to examine
the ability of normal commercial AE instrumentation to distinguish and
estimate the difference in lengths of cracks in metal samples in a
laboratory environment.
III. EXPERIMENTAL APPROACH:
a) Acoustic Emission Equipment
The existing commercial AE equipment was a 32 channel SPARTAN AE
system (CPM software) manufactured by Physical Acoustics Corporation. The
operation of this equipment was characterized and checked for calibration
by use of an AE Simulator manufactured by Acoustic Emission Associates.
Using a true root-mean-square meter the parasitic noise of the AE
system was measured as a function of the system gain. It was found that
certain system gains have relatively high parasitic noise and thus give
poor sensitivity. To enhance the sensitivity, a system gain of 36 db was
chosen so that the parasitic noise was only a small part of the pream¬
plifier noise. The bandpass of the AE preamplifier (Physical Acoustics
Corporation) and the Spartan system was 200-400 KHz. The threshold was
selected to be a fixed value of 42 dB which was some 3 dB above the
background preamplifier noise at the gain of 60 dB. In one test the
threshold was set at 40 dB by mistake. Physical Acoustics u30 resonant
sensors (diameter about 0.375 in.) were used since they are reasonably
small with good sensitivity. All AE data was taken in the high speed mode
and post-test processed or analyzed by hand from hit listings.
b) Simulated AE Studies
Simulated AE was used to study wave propagation characteristics
and determine the propagation velocity in the specimens. The source of the
simulated AE was a standard Pentel pencil lead break (0.3mm dia, 2H
hardness). For these studies a 3.62 x 16.0 x 0.258 in. 2024-T351 aluminum
specimen was prepared with a 2.0 in. long by 0.12 in. wide (with rounded
ends) center through notch (perpendicular to the long direction). The
notch allowed lead breaks on its surface. Figure 1 shows the geometry of
lead break positions as well as the sensor locations used for various parts
of the research. For these studies vacuum grease was used as the AE sensor
1004
couplant, and the sensors were held in place by rubber bands. Since lead
breaks have approximately constant amplitude, to simulate the fact that
real AE signals vary in peak amplitude the location studies were made with
a number of different thresholds relative to the peak amplitude. In
addition to use of the AE system waveforms were also taken with two differ¬
ent Nlcolet transient recorders (digitization rates of 2 and 10 MHz with
8K memory) . The waveforms allowed detailed studies to determine proper
settings of the AE system Rearm Time (RTO) , Rise Time Time Out (RTTO) , and
Single Channel Event Time Out (SCETO) parameters. Based on tjrpical
waveforms, the values selected were RTTO - 20us, SCETO - lOOOus, and RTO
- 40 us.
c) Extraneous Noise Tests
Several approaches were taken to deal with extraneous AE noise.
First, to eliminate the potential for rotation of the specimen relative to
the grips a symmetric center crack specimen was used. Second, smooth faced
hydraulic grips were used which apply the transverse load prior to the
axial loads. Third, an uncracked alumimim sample (made with grip marks
left by the precracking grips) was cycled at the typical crosshead rates
used to test the cracked samples. Two or three sensors were mounted with
rubber bands and vacuum grease or by adhesive bonding on the sample during
these tests. One sensor was placed at the center of the sample for some
tests along with two others near the grip edges for all tests. The load
to the sample was monitored by the AE System parametrlcs during these
tests. Table 1 shows the number of events for each increasing and
decreasing load cycle. Examination of the AE hit listings showed that the
few extraneous noise events generated hit one of the sensors near the grips
first. This result validated that use of guard sensors would allow identi¬
fication of any extraneous noise events. Fourth, the sensor array selected
for the cracked samples included two guard sensors in addition to the four
sensors placed around the crack region (see figure 1) .
d) Center Cracked Samples With Load Cycling
The 16.0 X 3.61 x 0.258 in. test samples were prepared from 2024-
T351 aluminum plate stock. A 0.50 in. long through slot in the 3.61 in.
direction was electron discharge machined (EDM) in the center of each
sample. The approximately 0.015 in. wide slot was extended by fatiguing
100-5
in an MTS machine at 10 Hz using a load shedding program so that Kmax
remained near the AE test level of 10 ksi/\in with R-0.01. During the last
0.020 in. of crack growth at each tip the maximum and minimum loads were
kept at the value used for the AE monitored fatigue cycles , Table 2 gives
the crack lengths of each sample at the conclusion of precracking. Since
the starter notch was the same for each crack, the three samples provided
three different lengths over which closure related AE could be generated
as well as three different crack tip positions. The flat precracks
Indicated that the cracks were grown in plane strain. Table 3 gives the
yield strength, and fracture toughness of typical 2024-T351 material
[Damage Tolerant Design Handbook, 1983]. The same reference also has data
which predicts that the crack growth rate at the end of precracking should
be between 0.1-0, 2 urn per cycle. The crack growth rate was measured at
about 0.4 urn per cycle during the last approximately 0.020 in. of growth.
For the AE tests the specimens were moved to a 100,000 lb capacity
Instron machine, which is much quieter than the hydraulic MTS machine with
servovalve attached directly to the actuator. Since it was impossible to
move the sample and grip it with exactly the same aHgnment in the new
machine, each sample in the Instron was cycled an additional 630 cycles at
a rate of about 0.4 Hz at the maximum and minimum loads used at the end
of the crack sharpening procedure. This procedure grew each crack tip an
additional nominal 0.005 in. and resulted in the crack tip adjusting to the
slightly different alignment.
The AE test procedure consisted of 6 to 12 cycles at the same load
levels used in the resharpening procedure. The crosshead rates and the
approximate times for one cycle for these tests are given in table 2.
Following these cycles an additional cycle with an overload above the
previous peak load level was Imposed. The load factors for the overloads
are given in table 2. The purpose of the overload was to generate a number
of events located at the crack tips as the plastic zone size was increased.
During the AE tests, six AE sensors were used as shown in figure 1. The
sensors on either side of the crack were the primary monitoring sensors,
while the two sensors near the hydraulic grips were used to identify any
AE signals which originated outside the crack region. The use of edge
mounted sensors in some tests versus surface mounted sensors was based on
100-6
an earlier observation that in rod type samples sensors mounted on a
specimen face perpendicular to the major stresses released at the AE source
gave better location results [Hamstad. et al., 1986]. Also, the edge
sensors were more sensitive by about 4-6 dB in the wave propagation
studies. The sensors were bonded in place using an epoxy mix (Cole Farmer'
#8778). After sensor installation lead breaks, about O.S in from the
center of each sensor, were used to check sensitivity. A signal propor¬
tional to the load cell output was connected to the parametrics input of
the AE system so that the stress intensity level for each AE event was
recorded in the data set,
IV. RESULTS AND DISCUSSION OF RESULTS:
a) Simulated AE Studies
The velocity of the first arrival wave packet (first negative
peak) was measured to be 214,000 in/s. This value corresponds to the in¬
plane extenslonal mode [Goman, 1990] . Using this velocity the expected
differences in arrival times,, delta T, for lead breaks at positions 1, 2
and 3 (see figure 1) in the notch were calculated. Table 4 shows these
theory results as well as the experimental values based on lead breaks for
one array of two sensors (Indicated in figure 1). The mean delta T values
as well as the maximum and minimum values for four to six lead breaks are
given as a function of the approximate difference measured in decibels
between the peak of the AE signal and the threshold, delta dB, There are
three main observations to be made from this data. First, the difference
between the experimental values and the calculated value tends to increase
as delta dB decreases. This means that the effective propagation velocity
is decreasing as the fixed threshold, which detects the arrival of the AE
signal, moves further into the analog signal. Second, the maximum and
minimum values of delta T also show a wider range as delta dB decreases.
This result implies that there is more variability in exactly which
positive or negative half-cycle crosses the AE system threshold. Third,
the mean value for the symmetric (with respect to both specimen geometry
and sensor locations) lead break position (position 1) does not change much
with decreasing delta dB. Results for the other sensor arrays were similar
except that nearer positions gave delta T values closer to calculated
100-7
values for a given delta dB, and as already mentioned edge sensor positions
experienced a 4-6 dB higher amplitude signal than surface mounted sensors.
The above results are clearly elucidated by the AE waveforms for lead
break signals. Figure 2 shows typical waveforms at two sensors for a non
symmetrical lead break. Since the AE system has a fixed threshold that
triggers either on positive or negative signals, the AE system is triggered
not on the first deviation of the signal from the background noise but
instead at some point within the first wave packet, as the threshold is
raised. Since for a non symmetric lead break location the AE signals at
the different sensors are not in phase, and have different amplitudes and
envelopes, differences between the measured delta T values and the
calculated theoretical values arise. For this experiment if the two
channels trigger out of phase by one-half cycle, the error is about 1.5 us
and a full cycle gives an error of about 3 us. These ^errors are much
larger chan the + one clock-tick (0.25 us) error of the system. Since the
individual cycles of the waveform both rise and fall within the first
packet, even greater errors can occur, leading to differences greater than
3 us between calculated and experimental values. Since a symmetric lead
4
break generates in-phase and similar amplitudes at the sensors, the
experimental values for delta T tend to be close Co the calculated values,
but the values can still vary by up to a cycle due to differences in
sensitivity in the two AE channels. One other important observation from
the transient recorder records can be seen in table 5. Here the calculated
theoretical delta T value is compared with Che delta T values which can be
obtained from the two channel transient recorder by making in phase
measurements of the delta T value f.'or a non symmetric lead break. Again,
it is seen that the effective propagation velocity increases even for
inphase measurements as the triggering point is moved into the waveform.
The implications of the above results for the crack-tip location phase
of this research are several-fold. First, only at delta dB's (between peak
and threshold) greater than say 18 dB can we expect the AE measurement
system to give reasonable values for delta T values. Even in these cases
sensitivity differences between channels or the known lack of AE source
radiation symmetry (Scruby, 1985] can lead to errors on the order of 1.5
us to 3 us or more. These errors when translated to distance can lead to
100-8
significant deviations in crack tip location considering the 0.214 in/us
velocit)’. Second, the delta T values may need correctipn by use of a
variable velocity for different delta dB values. Third, transient recorder
waveforms gathered in the crack tip experiments can be expected to provide
more reasonable results because in-phase delta T values can be determined.
But, again an error in the crack tip location will result unless a variable
velocity of propagation is used. The above results are all related to the
fact that the peak amplitude of an AE signal is about 40-55 dB greater than
the peak amplitude of the first half cycle deviation from background noise
in the current experiments. This fact means that for all of the AE signals
from the crack tip sources, wavefonns will have the early part of the AE
signal burled in the preamplifier noise because the peak amplitudes are not
large enough. Hence the majority of delta T's which are determined will
come' from points somewhere within the AE waveform and’ not from the first
arrival of the signal.
b) Center Cracked Samples With Load Cycling
The crack tip AE source, which will generate AE at peak loads or
during an overload, is expected to be inclusion particle fracture in the
plastic zone in 2024-T251 aluminum [Scala and Cousland, 1983]. Crack
growth during fatigue is not expected to be a source at the sensitivities
used in these experiments because the micro processes involved take too
long to generate significant AE. Since relatively few events occur at the
crack tip during fatigue cycling to a fixed level [McBride and Maclachan,
1982, Scala and Cousland, 1985], overload AE was studied to accomplish the
crack tip location objective of this work.
Standard AE location algorithms for a four sensor rectangular array can
be used to examine the capability of AE technology to distinguish the three
different crack lengths used in these experiments. But, due to the errors
in delta T values which are expected as a result of out of phase arrival
times, such an approach is not expected to lead to reliable results.
Figure 3 shows the result is as expected. Hence, the approach was changed
to analyze the data in a more meaningful fashion. Using event listings,
hand calculations were made of delta T values that would locate a crack tip
event. The procedure was to first use the sensors symmetrically located
above and below each crack tip to calculate delta T values. For crack tip
100-9
events these values should be nearly zero assuming symmetrical source
radiation. Next the delta T value related to the location along the crack
length was determined using the two sensors above (below) the crack. Table
6 grves the raw data from an event listing, and the sample calculations for
an event with a relatively large delta dB between the threshold and the
peak amplitude. Table 7 shows this result as well as data more representa¬
tive of the fact that most of the crack tip events had delta dB values
considerably smaller than the sample event in table 6.
The relative ability to locate crack tips by use of AE data analyzed
with the above procedure is shown in table 7. This table gives delta T
values based on the following approaches: i) calculated theoretical values
using the measured extensional velocity; ii) measured values using surface
lead breaks at the crack tips for two different thresholds; iii) AE system
measurements (analyzed as in table 6) from overload events; and iv)
measurements taken from transient recorder records (10 MHz) from events
during overloads (table 7b). If the delta dB level is large enough, the
three different lengths can be distinguished by the commercial AE system
data, But, since in general the delta T values are too large, the crack
length values calculated from these delta T's would tend to be larger than
the actual ones. There is an additional very significant result to be
gained from this table. The generally wide deviations from the expected
value of near zero for delta Tja-is and delta Tig-is is not expected based on
the wave propagation results. The likely explanation is the previously
noted lack of uniform radiation of sound energy from the AE sources. This
condition implies an additional source of error for fixed threshold
approaches .
On the other hand, table 7b shows that the transient recorder measured
delta T values can be much closer to the calculated theory values. The
values listed in this table were determined by measuring the delta T
between the first positive or negative peak (positive peaks in both
channels or vice versa) of the AE signal above the noise. As can be seen
in the table, there are some difficulties with the current transient
recorder approach as well. The main problem is that it is difficult to
determine which are the two inphase peaks from the two AE sensors. Since
the real AE signals have nonsymmetrical radiation of energy, the determina-
100-10
tion is not trivial, and the above approach may not have determined the
inphase points. Some preliminary data, based on analysis of a few hard
copy waveforms from the long crack case, indicates that in some cases a
small precursor wave can lead to errors in determining the inphase arrivals
using the above approach. The limited studies here did not fully identify
a standard approach to this inphase determination. It may be easier to use
a different type of AE sensor that will simplify the problem (Gorman,
1990]. But such an approach must maintain sensitivity and applicability
to real structures.
In addition to AE from the crack tip at peak loads, a second period of
AE occurred on 'ch fatigue cycle. Due to the low R ratio (R - 0.01) crack
closure was present for the lower loads of each cycle. Crack closure did
not generate significant AE at the sensitivities used. But, the "unclos¬
ure" deformation generated significant amounts of AE for each fatigue
cycle. The source mechanism of this crack opening AE has not been a source
of detailed study [Guozhi, 1985). A potential source is the relatively
brittle fracture of the cold welds, formed when the two crack surfaces come
together as the load is reduced in the fatigue cycle. Table 8 gives the
approximate number of "unclosure" events for each crack length as a
function of rest time at minimum load. The approximate stress Intensity
range over which this AE was generated is also listed in this table. A
number of significant observations can be made. First, the number of
"unclosure" events increases significantly with increased rest time.
Second, the number of "unclosure" events changes small amounts (for a fixed
rest time) with increasing length of precrack (total crack length minus
starter notch length) . Another related observation is that the number of
crack tip events per cycle was relatively few compared to the number of
"unclosure" events. Typically the number of crack tip events was zero to
three events per cycle. Attempts to apply the AE system source location
approach used for crack tip events (see table 6) were plagued by the same
errors in delta T values which are present in the crack tip events.
Further, since the location of these "unclosure" events was not known,
there is no standard to compare the results as was done for the crack tip
events. Using transient recorder (10 MHz) records, delta T values were
determined from in phase positions (assumed to be first same sign cycle
1(X)-11
peaks) for the "uhclosure" events. Table 9 shows these results for the
long and short cracks along with the approximate load at which the events
occurred. The results are not easy to interpret. In some cases the region
of the starter notch corresponds to the delta T value . It may be that more
sophisticated approaches are needed to find the true inphase positions.
The limited number of waveforms examined here did not lead to such an
approach. This potential unzipping of crack closure seems to represent an
additional opportunity for crack detection and length characterization.
There are two reasons why this source of AE should not be neglected.
First, this source is much more numerous than crack tip events and thus
provides more data. Second, this source is much more likely to continue
to operate under spectrtun loading which occurs in real applications. Some
key questions still to be answered about this source include, over what
length of the crack does it occur? Does this source occur in very ductile
materials which have been observed to not emit AE with crack growth? What
are the angular radiation properties of this source? And what exactly is
the source mechanism?
V. STATUS OF CBACK-TIP LOCATION BY AE:
This research shows clearly that AE technology as practiced with
commercial instrumentation currently available cannot accurately determine
the crack length even under relatively ideal conditions (small sample,
laboratory environment, and ideal sensor locations). The key factor that
leads to errors is the measurement of AE signal arrival times by a fixed
threshold.
The primary problem is the measurement technique and not the basic AE
waveform data which can provide reasonable results. Using displacement
type point contact sensors which allow identification of specific wave
modes it has been shown that techniques using waveform recorders can quite
accurately locate crack tips in relatively uniform laboratory size samples
[Scruby, 1985] . Unfortunately, these sensors do not have high sensitivity.
The approach of using waveform recorders has some difficulties. First,
waveform recorders are not designed to handle the event rates in typical
AE testing. If applied to AE instrumentation, current technology can
overcome this difficulty. Second, low amplitude AE sources can sometimes
100-12
pose difficulties in finding the inphase arrivals at different sensors due
to lack of symmetry in energy radiation from an acoustic source.
VI. CONCLUSIONS:
1) A center crack specimen, screw machine, and hydraulic grips
leads to elimination of extraneous noise in fatigue crack
propagation.
2) Commercial AE systems with fixed thresholds cannot in. general
accurately locate crack tips for most of the cra^k tip events
in this case.
3) Use of waveform recorder* to determine the arrival time with
the same phase at each sensor can give reasonably accurate
crack tip locations.
4) "Unclosure" AE dominates crack tip generated AE,
5) "Unclosure" AE may offer a,n additional approach to crack
detection and length characterization.
VII. RECOMMENDATIONS:
1) Develop algorithms to rapidly locate the earliest possible in
phase arrival times of digital AE wavefenas for an array of AE
sensors even in t?ie presence of precursor yayps or low. signal
to noise ratios.
2) Examine the use of ultrasonic type sensors or other sensors to
make such an algorithm simpler (Gorman, 1990]
3) Study the waveform approach for larger te.et samples. First,
for uniform tnickness samples with a range of thicknesses and
later for iaraplos with nonuniform tliicknesses and other
appropriate geometry. Determine the size of areas .vhi<;h can be
effectively monitored. Real AE sources may be nsces.'oary since
non-symmetry of source radiation must be included in the study.
4) investigate in mere detail the "unclosure" AE source in both
aluminvra and more ductile materials, Also include in the study
spectrum fatigue at realistic loading rates.
5) Document the- sehixtivity improvemente available for AE sensors
relative to the .’’iplitude of AE sources of interest..
iQ0*i3
REFERENCES
Damage Tolerant Design Handbook, Vol, 3, HCIC-HB-OIR, Metals and Ceramics
Information Center, Battelle Columbus Laboratories, Columbus Ohio , 1983, p.
7.5-1, 7.5-111. .
Gorman, M. R. , Plate Wave Acoustic Emission, Submitted for publication to
the Journal of the Acoustical Society of America, 1990,
Guozhl, Lu, Fatigue Crack Closure Study, Jouxmal of Acoustic Emission, Vol.
2, No, 2/3, 1985, pp. S203-S206.
Hamstad, M. A. , R.D. Young and P. M. Thompson, Acoustic Emission Source
Location During Four-Point Bend Tests in Alumina,. Progress in Acoustic
Emission III, Japanese Society of NDI, 1986,:,pp, 26-33.
McBride:, S, L. and J. W. Maclachan, Acoustic Emission Due to Crack Growth,
Crack Face Rubbing, and Structural Noise in the CC-130 Hercules Aircraft.,
J. of Acoustic Emission, 3 January 1984, pp. 1-10.
Seals, C. H. and S.'McK.. Cousland, Acoustic Emission during Fatigue Crack
Propagation in Aluminum Alloys 2024 and 2124, Material Science and
Engineering, Vol. 161, 1983, pp. ^l-2i8.
Seals, C. M. Snd S. McK. Cousland, Acoustic Emission During Fatigue of
Altiffiih’jm Alloy 2024; The Effect of an Overload, Materials Science and
Engineering, Vcl. 76, 1985, pp. 83-88.
Sciruby, C. B. , Quantitative Acoustic Emission Techniques, in Research
Techniques in Nondestructive Testing. Vol. Vlll, Edited by R. S. Sharpe,
Academic Press, London, 1985, pp. 141-210.
Sendeckyj , G. P. and C. A. Paul, Some Smart Structures Concepts, Fiber Optic
Smart Structures Snd Skins II. SPIE Proceedings, Vpl, il70, 1989, pp. 2-10.
100-14
Table 1 Extraneous AE during test with, dtpmy sample (loading
between 200 lbs and 14,000 lbs)i
Total Events
Increasing
Decreasing
Condition
Load
Load
a. Newly installed sample:
first cycle
12
9
second cycle
4
2
third cycle
6
2
b. Sample reinstalled:
first cycle
19
19
second cycle
2
9
Table 2 Conditions for center
cracked specimens
Crack length
Crosshead
rate , in/min
Overload
Specimen 2a (in)
(cycle time, min)
factor
Short Crack 0 . 741
0.005 (21)
1,18
Medium Crack 1,755
6.002 (18)
1.53
Long Crack 2.770
0.002 (7)
1.09
Table 3 Mechanical orooerties:
2024 - T351 aluminum.
Yield strength: 54.5 ksi
Plane strain fracture toughness: 35 ksi ^ in
Table 4 Typical lead break wave propagation times for edge
sensors at nominal 3 in. position (notched sample).
Lead break
position and
theoretical
delta T, us
Delta T, Mean and range (us)
52dB
30dB
18dB
6dB
1
-0.4
0.2
0.4
0.7
(0)
(0.3 to
0.5)
(0.2 to
0.3)
(0.3 to 0.6)
(0.2 to 2.5)
2
3.0
2.6
2.2
4.8
(2.4)
(2.6 to
4.0)
(2.4 to
2.7)
(1.0 to 2.8)
(-103 to 4.8)
3
5.2
5.2
4.4
28.5
(4.4)
(4.4 to
6.2)
(4,7 to
6.2)
(3.0 to 5.0)
(-96 to 28)
100-15
Table 5 Change in delta T for inphase measurements for sur¬
face lead breaks ac crack tip (long crack sample) with edge sensors and
transient recorder measurement. Using times at peaks of half-cycle of
waveform.
delta T, us
First negative
8.8
First positive
9.3
Second negative
9.6
Second positive
9.9
Third negative
10.1
Third positive
10.5
Fourth, negative
10.6
Fourth positive
10.8
Fifth negative
11.3
Fifth positive
10.4
Theory (9.4 us)
Table 6 Sample calculation of delta T values using rectangular array of
sensors, for an overload event (long crack sample with delta dB
- 26dB)
13 _ 15
18 19
Theory ,
Arrival absolute
Channel
time, us
delta T Calculation
value ,
18
87.0
delta T^g_i3
- 87.0-89.0 - -2.0 us
(0)
13
89.0
delta Tj^9_^5
- 95.0-95.0 - 0 us
(0)
15
95.0
delta Ti3_i5
- 89.0-95.0 - -6.0 us
(9.4)
19
95.0
delta Ti8_i9
- 87.0-95.0 - -8.0 us
(9.4)
Table 7a Some typical delta T values for overload events using data from
AE measurement system (*largest event)
Crack length Delta T, us Delta dB
Short with
13-13
-1.5
19-15
2.6
13-15
7.3
18-19
3.2
2
edge sensors.
4.0
-1.2
2.0
7.2
4
theory delta T
0.6
-3.2
-6.0
-2.2
8
-2.8 us
1.0
-2.0
-4.6
-1.6
12
-2.2
3.0
5.2
0.0
16
0
0
-2.7
-2.7
18*
Tip lead break
2.0
0.6
3.0
4.4
14
Tip lead break
2.0
0.4
-3.0
-1.4
15
Tip lead break
0.3
0.6
3.0
2.7
45
Tip lead break
1.8
0.2
-4.6
-3.0
45
100-16
Medium with
8.0
0.4
7.0
14.6
3
top sensors,
0.2
3.2
-4.6
^7.0
8
theory delta T
-1.0
0.7
-7.7
-9.4
14.
-5.5 us
10.6
0.5
6.0
15.5
19
2.0
0.5
6.0
7.5
24
0.0
0.6
-7.2
-7.8
24
0.2
0.3
7.3
7.2
36*
Tip lead break
1.4
0.4
6.0
7.0
48
Tip lead break
0.8
0.4
-7.6
-7.2
50
Long with
5.0
0
12.2
7.2
6
edge sensors,
37.8
5.2
-11.6
21.0
9
theory delta T
5;7
5.5
9.8
10.0
11
-9.4 us
2.0
8.5
-8.5
-15.0
14
1.3
0
-11.0
-9.7
15
-2.0
0
-6.0
-8.0
26
-2.4
0
12.0
9.8
30*
Tip lead break
1.3
0
-11.0
-9.7
15
Tip lead break
1.4
-0.3
10.3
12.0
15
Tip lead break
0
0
-10.2
-10.2
46
Tip lead break
-0.2
00
o
1
9.8
10.8
44
Table 7b Typical delta T values from transient recorder for
overload events
Crack length
Delta T, us
Peak amplitude*
15-13
mv
Short with
-9.5
35 to 46
edge sensors.
-3.1
66 to 70
theory delta T
-2.6
65 to 81
—2.8 us
2.0
162
-4.0
185 to 193
2.2
188 to 229
-2.8
185 to 234
Tip lead break
-2.7
2,700
Tip lead break
2.8
2,600
Long with
edge sensors
-9.1
33 to 39
theory delta T
-6.0
165 to 170
- 9.4 us
-2.0
46 to 50
10.1
170 to 240
Tip lead break
-8.5
2,800
Tip lead break
9.0
2,900
*Preamplifier noise typically 16-20 mv pp.
100-17
Table 8 Typical number of crack opening ("unclosure") events and
approximate load range of AE source operation versus rest time .
Number of events/Rest time (min)
Load ranee
Kj, Ksi /Tin
Short Crack
Medium Crack
Long Crack
0.7 to 2.8 86/1.4
0.9 to 3.0 82/0.8
1.0 to 3.6 60/1.1
161/58.6
170/63 397/988
105/110.5
Table 9 Delta T location information for crack opening
("unclosure") events from transient recorder*
Short Crack
Long Crack
Delta T
13-15 us
Peak ampli¬
tude, mv
Load, lbs
-3.8
82
to
99
695
1.8
72
to
86
800
-2.4
103
to
120
900
-2.0
112
to
141
1210
-2.2
188
to
190
1360
2.0
205
to
229
1720
1.8
196
to
239
2360
2.2
286
to
327
2580
-1.2
30
to
60
200
2.0
128
to
226
315
-2.4
186
to
219
500
-0.8
84
to
94
850
-0.9
>320, saturated
1105
-2.1
176
to
199
1325
-2.4
275
to
309
1435
1.8
>320, saturated
2210
* Theoretical delta T values 1.8, 2.8, 9.4 us for starter
notch, short crack tip, and long crack tip respectively
100-18
q.
Fleure 1 Location of centers of sensors and notch leadbreak positions,
symmetrical about lines through notch or crack and perpendicu¬
lar to crack. Key: x Lead break positions in notch, 0 propaga¬
tion study sensor position, D test sensor positions, ^ guard
sensor positions, and ^ centerline of crack. Dimensions in
inches, not to scale, one quadrant of sample.
Ch 1
Ch 2
Figure 2 Typical waveforms for lead break at crack tip of long crack
sample with edge sensors (Channel 1 near sensor, Channel 2 far
sensor, at 40 dB gain, vertical scale Iv/div, horizontal scale
lOus/div)
100-19
90
30
10
-10
-30
-50
Figure 3 AE measurement system location of crack tip events from
overload cycle for standard system algorithm for rectangular
array on medium crack sample. Key: Event locations Wk ;
coordinates of sensors 0,0; 13,0; 0,14; 13,14; units of axis in
us .
MEOIUn CRACK 13 UP
08/16/90 10130106
Acknowledgements
I would like to thank the Air Force System Command and the Air Force
Office of Scientific Research for sponsorship of this research. In addition
I also want to thank Universal Energy Systems for the excellent way in which
they handled the administrative aspects of the program.
The experience of this research was very interesting as well as
challenging. I want to particularly thank Dr George P. Sendeckyj , for his
help in obtaining all the necessary equipment and specimens for completion
of the summer research plans. All the help that David L. Hart gave me with
the acoustic emission equipment was greatly appreciated. Finally, I want
to thank Harold Stalnaker, Larry Bates and Don Cook for carrying out the
testing of my samples, and I also to thank Lt Joe Storr for his interest in
making sure my needs were met. Also, the excellent t3rping of Melinda Pizzo
is greatly appreciated.
100-20
1990 USAF-UES SUMMER FACULTY RESEARCH FROORAM/
GRADUATE STUDENT RESEARCH PROGRAM
Sponsored by the
AIR FORCE OFFICE OF SCIENTIFIC RESEARCH
Conducted by the
Universal Energy systems^ Ino.
FINAL REPORT
H^ Design Based on Loop Transfer Recovery and Loop Shaping
Prepared by:
Academic Rank:
Department and
University:
Research Location:
USAF Researcher:
Date :
Contract No:
Chin S. Hsu, Ph.D. Jenny L. Rawson
Associate Professor Graduate Student
Department of Electrical Engineering and
Computer Science
Washington State University
Flight Dynamics Laboratory
WRDC/FIGC
Wright-Patterson AFB, OH 45433
Siva S. Banda, Ph.D.
17 Aug 90
F49620-88-C-0053
Design Based on Looy Transfer Recovery and Loop Shabiaa
by
Chin S. Hsu
Jenny L. Rawson
hismsi
This report addresses the issue of loop transfer recovery and
loop shaping when an output feedback controller is used. A method of
selecting the design parameters to achieve asymptotic loop transfer
recovery is presented. It is shown that the problem of approximate loop
transfer recovery is equivalent to that of state feedback design. A
new design procedure is also presented.
101-2
Ackncwledaemehta
We wish to thank the Air Force Systems Command and the Air Force
Office of Scientific Research for sponsorship of this research.
Universal Energy Systems must be mentioned for their concern, and help to
us in all admin^trative and directional aspects of this program.
Our experience was rewarding and enriching. John Bowlus provided
us with support, encouragement, and a truly enjoyable working
atmosphere. The help of Dr. Siva Banda and Dr. Hsi-Han Yeh was
invaluable for furnishing us with much-needed technical guidance. Their
concern and keen interest in our research progress was the key factor
which rendered our research effort into a pleasant and fruitful task.
We also wish to thank Capt. Sharon Helse for helping to make our work
easier and fun.
101-3
I . Introduction :
Robust multivariable control has been a main thrust of modem
feedback design. Practical control systems must be designed to cope
with various system uncertainties. One particular concern of robustness
is the stability measure in terms of the gain and phase margins which
are determined by the system's loop transfer functions. An effective
approach to design a controller for robust control is the Linear
Quadratic Gauss lan/Loop Transfer Recovery (LQG/LTR) method. The LQG/LTR
output feedback design provides a means of recovering the guaranteed
margins which are an inherent feature of state feedback design via the
Linear Quadratic Regulator (LQR) .
The most recent robust control design is based on the
optimization method. The method ensures that the closed-loop system
is internally stable and the norm of the closed- loop transfer
function matrix is less than a prespecified scalar index. Just like the
LQG/LTR method, suboptimal controllers can be obtained by solving two
Riccati equations.
A significant extension of the design has been made at
WRDC/FIGC. This is a powerful and new control design procedure using
mixed Ha and optimization.
My research interests have been in the area of linear control
systems with emphasis on developing numerical algorithms for industrial
applications. My previous research on model reduction, system
identification and optimal control, which serves as a general background
101-4
for robust control theory, contributed to my assignment to the Control
D3mamics Branch, jPlight Control Division.
II. OBJECTIVES OF THE RESEARCH EFFORT:
The objective of this summer research is to explore new design
methodologies in robust multivariable control. Motivated by the new
H2/H„ controller design procedure, we focused our efforts on extending
design by incorporating loop transfer recovery and loop shaping. The
goal of developing this new design approach, H^LTR, is to provide
control engineers with a procedure for designing practical robust
control laws for aerospace applications in the presence of unwanted
external disturbance.
Owing to the facts that suboptimal controllers are observer-
based and obtained from the solution of two Riccati equations, it is
natural to investigate the possibility of designing compensators
using the combination of an H„-state feedback regulator and a full-order
observer. The success of this new observer-based design will yield
an effective method of developing reduced-order controllers with the
use of reduced- order observers.
III. H^LTR: A LOOP SHAPING METHOD FOR H„ OUTPUT FEEDBACK
JONPENSATOR DESIGN:
1. Motivation.
State feedback with an constraint as presented in [3] gives a
static controller from the solution of a single Riccati equation. The
objectives are to provide internal stability and to bound the norm of
the transfer function between a deterministic disturbance input and a
101-5
designated output. Additional benefits are a bound on the sensitivity
function and guarantees on the gain and phase margins at the input to
the plant that are even better than those given by linear quadratic
regulators (LQR) .
These benefits may not be maintained when output feedback is used.
In the Linear Quadratic Gausslan/Loop Transfer Recovery method (LQG/LTR)
[8, 9], an adjustable parameter in the filter Riccati equation is used
to asymptotically recover the state feedback loop transfer function when
output feedback is used. This allows the stability margins to be
improved. In this paper, we investigate a method for asymptotic
recovery of the state feedback loop transfer function when output
feedback is used (H^LTR) . The goal is to regain the good properties of
H_ state feedback.
40
One disadvantage of asymptotic LQG/LTR is that the feedback gains
can become unacceptably high; the same problem can occur in H^/LTR. To
avoid this, approximate H^/LTR can be used. This paper includes a
method that is based on a Riccati equation which solves a dual state
feedback problem.
In the following, section 2 includes a description of the feedback
system, formal statements of the H^LTR and loop shaping problems, and a
review of the state and output feedback methods. asymptotic and
approximate LTR methods are presented in section 3. In that section, it
is also shown that some asymptotic loop shaping can be done in the state
feedback case . Proofs for the theorems and lemmas can be found in ( 7 ] .
101-6
A. B, C
Di , Djt
Vi, Vs
^3gs
Rico*
Q.
positive numbers.
h, m, m, q, p, h-dimensionaX vectors,
nxn, iixm, mxn real -valued matrices,
nxq, mxq real-valued matrices.
DiDi'i CsDa^.
pxn, pxm real-valued matrices.
EsJEs«.
nxn symmetric, positive semi-definite matrices.
Aq, Bg, Cg nxn, nxm, mxn real-valued matrices.
G(s), Gg(s) {A, B. C), {A^., Bj., Cg).
Ls(s), Lo<s) {A, B, Cj.), Gg(s)G(s),.
RgCs), RqCs) I - LjjXs). I - L^Cs).
E(s) Lg(fl)-Lo(s).
A, M(s) A + {A - BgC, B, C^).
2. Pr«liainari«8 and Problaa Statamant.
The plant to be controlled is described by the following
equations .
x(t) - Ax(t) + Bu(t) -f Diw(t), x(0) - xo.
y(t) - Cx(t) + D2w(t) (1)
2(t) - Ei„x(t) + E2„'l(t)
A goal of design is to bound the norm between the deterministic
disturbance w(t) and the output z(t) while internally stabilizing the
closed- loop system. To do this, the following assumptions will be made.
101-7
Assumptions :
1. {A, B) and {A, Pi) .are stabilizable pai^fs;
2, {A, C), and {A, Ei„): are detectable pairs ;
3. Ei/E2^ -P;
4, V2 - psPa^ > P. Ra* 22oo^Ea«, > 0.
Two problems are addressed Ih this paper. The first is to design
ah output feedback compensator to recover the loop transfer function as
designed in a state feedback problem'i The second is to shape the state
feedback loop transfer function. These are both stated formally below.
Find design parameters Pi and P2 such that output feedback
as3rmptocically gives the same loop transfer function at the input as
does state feedback.
S.;. Loop Shaping.
Find design parametars Et„ and E2i„ such that state feedback
gives a desired loop transfer function.
Solutions to the.^e two problems are presented in Section 3. The
remainder of this section is devoted to a review of the standard state
and output feedback design methods.
State Feedback. (1, 3]
In this case, C - I and D2 ■ 0. Static state feedback is used,
with one Riccati equation to be solved. This equation is:
0 - Y„A + a'^Y^ + Y„(7'2vi - BR2„’ V)Y^ + Ri„. (2)
The control law is then u(t) - CgX(t) where C^, - -Rg^'^B^Y^. Let Hg(s)
be the transfer function between w and z. The closed- loop system is
101-8
stable, and l|Hg(s) ||-^, < 7 iff 5: 0 and A + (7'^V,i - BR2*'V)Y|, is
stable.
Transfer functions of interest are the loop transfer function
Lg(s) - Cg*(s)B and the return difference matrix Rg(s) - I - Lg(s;,
where #(s) - (si - A)*^. It can be shown that
- R2« + B'^(-rjw - A‘^)'^Ri„(jw - A)'‘iB
+ 7‘V(-jw - A’^)-^Y„ViY„(d« - K)-h (3)
This can be used to determine guaranteed gain and phase margins much as
is done with ,LQR (4* 8, 9]. Here there is the added benefit of an extra
positive semi-definite term on the right hand side of equation (3) to
give even better margins than those of LQR [2],
H<o,, . li, 6)
A dynamic compensator is used and two Riccati equations must be
solved when the states are hot available, the design equations are;
0 - Y^A + A'^Yi^ + Y^(7'2Vi - BRj^," V}Y„ + Rx„
0 - AQ + QA^ + Q(7’^Ri„ - c'^V2’^C)Q, + Vx (4)
And the compensator is:
;tj,(t) - A^Xg(c) B^yCt)
u(t) - C^Xj,(t) (5)
where ,
Ac - A - B^C + BC^ + 7-'vxY„
B^ - (I - 7'Voo)'^Qc'^V2*^ (6)
Cc - -R2»’^bTy„.
Note that is the same as for state feedback. Figure 1 is a block
diagram of the closed- loop system.
101-9
Figure l. closed- Loop Control System with Compensator.
101-10
Let Hq(s) be the transfer function between w and Z; There will be
a stabilizing solution such that |1Hq(s)||^ < 7 iff Q, 2: 0,
2
pCQ^a,) <7 > (pO denoting the spectral radius of -a matrix), and both
A + (7‘^Vi - BR2„"^b''‘)Y„ and A + Q(7'^Ri„ - c'''V2‘^C) are stable.
As will be shown in the next section, there may not be any
guarantees on the gain and phase margins at the input to the plant.
3. H^LTR Design Methodology
3.1 LTR Conditions
Before presenting the actual design theory, we would like to
examine some conditions for exact LTR. This is done in the next lemma
and two theorems.
Lemitia 1: Define «^(s) - (si - A^,)*^, Lq(s) - Cjj»g(s)Bj,C«(s)B (the
loop transfer function from u" to u' ) and Rq(s) - I - Lq(s). Then,
Ro’^(-j«)R2«Ro(j«) - Rj?(-j«)R2«Rs(j«) - Ls'r(-j«)R2„L3(j«)
+ Lj(-j«)R2„Lo(j«) (7).
Proof. Follows by using the definitions of Rg(s), Rq(s), Lg(s) and
I.o(s). ■
This lemma suggests that output feedback may' or may not give
phase and gain margins which are as good as those guaranteed for state
feedback. If they are not, then the use of LTR is indicated.
Theorem 1: (H<„/LTR Condition)
With Vi - DiDJ - BVB^, V > 0, then Lq(s) - Lg(s) iff
B^[I + C$(s)B^,]'^C$(s)B - B. (8)
101-11
Note 1:
+ C$(s)Bg]‘^C$(s)B - B «» [I + C*(s)Bj,]'^B - 0 o B - 6 .
This indicates that B^, will have to be chosen carefully in order for
exact recovery to occur.
Theorem 2: (Mismatch Condition)
Again let Vi - DiDJ - BVB*'', V > 0. Define E(s) - Lg(s) - Lq(s)
and M(s) - Cj,(sl - A + Bj,C - 7‘2bVB‘'‘yJ*^B. Then
E(s) - M(s)[I - M(s)]*^[I - (C^ + 7-2vB'rYJ»(s)B] , (9)
and
E(s) - 0 iff M(s) - 0.
■
Note 2: M(s) is the transfer function of the compensator with the
loop broken at XX in figure 1.
Note 3: The above lemma and two theorems are equivalent.
The following lemma gives an expression for the sensitivity
function at the plant input, and shows that with LTR, the sensitivity
2
for output feedback approaches that for state feedback. If - n 1,
then equation (3) can be used to show that ||Sg(s)||^ < 1. Thus, the
upper bound on ||Sq(s)||^ can be made arbitrarily close to 1 by decreasing
P-
Lemma 2: Let Sq(s) •• Rq'^(s), Sg(s) - Rg'^(s). Then,
Sjj(s) - [Rg(s) - 7’^M(s)VB'rY„$(s)B]*^[I - M(s)]. (10)
And, if M(s) ■ 0, then
S„(s) - 83(3).
The proof requires simple manipulations. ■
101-12
3.2 LTR Method
The weights for LTR are Vi - DiDi^ - BVB^, V > 0, and D2 - pD2,
/> > 6 where we require that D2D2^ > 0 and DiD2^ ■ 0. The design
equations become:
0 - Y„A + + Y„(7"2BVB'r - BR2„- V)Y, + Ri„ (11)
0 - AQ + Qa"^ + Q(7'^Rioo - p'^c'''(D2D2''')'^C)Q + BVB’’’ (12)
And the compensator is:
Ac " A - B^C + BC^ + 7’2viY„
Bq - (I - 7'Vco)'^p‘W(D2D2'^)*^ (13)
Cc - -R2«’^bTy„.
The next theorem shows that the state feedback loop transfer
function is recovered as p-»0.
Theorem 3: If G(s) - C#(s)B and Gj,(s) - Cj,#g(s)Bg, then
lim Gj,(s) - Cg*(s)BG'^(s)
p-*0
That is,
lim Lo(s) - lim Gj,(s)G(s) - C^,$(s)B - Lg(s). ■
p-*0 p-»0
Note 4: If for state feedback ||Hg(s)||„ < 7, then there exists a
po > 0 such that with output feedback ||Hq(s)||„ < 7 for all 0 < p < po.
3 . 3 Approximate LTR
As3raiptotic LTR is an iterative procedure, with adjustments of the
design parameter p alternating with solutions of the Riccati equations
and checks on the stability margins. To avoid this, an approximate LTR
procedure can be used instead, where the norm of the error E(s) is
constrained by an upper bound.
101-13
Theorem 4:
a) Approximate LTR can be achieved if for a frequency band of
interest, a(E(jw)) < a, a > 0, This is equivalent to requiring that
supcr(M(jw)) < S for some 5 > 0, where 5 is a function of o and
a(I-(C^,+7-2vB'rYJ*(jw)B).
b) Define A - A + 7*^ViY,o. If {A^ - C^CX, B^} is a detectable
pair and there exists a solution X 2: 0 to the Riccati equation
0 - AX + XA"^ + X(5'2c^,‘''c^, - c‘'‘g)X + BB'’’, (14)
then, sup(7(M(jw)) - 1|M(s)11„ < S. ■
Note 5: Approximate LTR is obtained by solving a second state
feedback problem,
3.4 Asymptotic Loop Shaping with State Feedback
In this section, we present a method for shaping the loop transfer
function when state feedback is used.
Suppose that the desired loop transfer function is -Cjj(sl - A)*^B.
To select the state feedback gain such that the actual loop transfer
function shape approaches this one and the closed- loop transfer function
satisfies the constraint, use the following weights:
Ejg, - Cjj, (A, Cp) a detectable pair;
®2oo “ E2a,^E2a, - I, Cp^E2a, " Oj
DxDJ - bvb''^, V > 0.
Then, the design equation becomes:
0 - Y^A + A^Y„ + Y^(7'^BVB'^ - m-2bb'^)Y„ + Cp\ (15)
And the feedback gain is - - /i’^B^Y^j.
101-14
U-s
2 2
Theorem 5: If y /fi > p(y) , then there is a solution > 0 to
( 15 ) . Furthermore ,
lim nC^(sl - A)’^B - -C„(sl - A)*^B.
^ c u
fi-*0 m
3.5 Discussion
The H^LTR procedure is analogous to LQG/LTR. In fact, it can
easily be shown that equations (8), (11), (12) and (13) reduce to those
for LQG/LTR [8] as y-^o. In addition, (2) and (8) are identical in form
to the "Kalman Inequality" and "Doyle-Stein Condition" respectively [9],
As with LQG/LTR, there is a dual procedure for H„/LTR at the
output of the plant. The H^LTR method can be extended to minimum
phase, non-square plants if the number of outputs exceeds the niunber of
inputs for recovery at the plant input, or the reverse for recovery at
the output.
Note that this procedure differs from McFarlane and Glover's
loop shaping [i] where the singular value plots of the open loop
transfer function are shaped i
3.6 Design Procedure.
The algorithm for loop shaping and H^/LTR is shown in the flow
chart of Figure 2.
101-15
101-16
Figure 2 . H^LTR Design with Loop Shaping
IV. DESIGN— A I^W APPROACH:
it is known that suboptimal controllers can be obtained by
solving two Riccatl equations and that the compensator is observer-based
[1, 9]. In this section, we propose a direct approach to design
using observer theoiry. This new approach has the advantage of
generating reduced- order controllers.
Consider the following linear system:
x(t) - Ax(t) + Biw(t) + B2u(t), x(0) - X6.
z(t) - Cix(t) + Dnw(t) + Di2«(t) (16)
y(t) - C2x(t) + D2iw(t)
Lemma 3: The closed-loop transfer function matrix (TFM) from w to
z is
Tgp(s) - Dll + (Cl - Di2K)(sI - a + B2K)*^Bi (17)
if state feedback is used; l.e. y(t) - x(t) and u(t) - -Kx(t). ■
Theorem 6: Suppose that an obseirver, with observer gain L, is
used with (16). It is of the form:
x^,(t) - (A - LC2)Xj,(t) + Ly(t) + B2u(t) (18)
u(t) - -Kx^(t)
then closed- loop TFM from w to z is
T(s) - T3p(s) + F(s)(sl - A + LC2)-^(Bi - LDzi)
where ,
F(s) - D12K + (Cl - Di2K)(sI - a + B2K)'^B2K (19)
101-17
Based on the foregoing result, an design procedure is as
follows :
1. Determine a state feedback gain K such that ||Tgp(s)||„ < 7.
This can be done by solving one Riccati equation [4] .
2. Determine an observer gain L such that ||T(s) - Tgp(s)||„ < S.
Note 6: The procedure in step 2 is dual to an Hj,,-state feedback
problem (i.e., filtering) , and as such, a Riccati equation can be
found to determine the observer gain L. Progress is being made to
demonstrate this design procedure.
Note 7: If, instead of a full-order observer (18), a reduced-
order observer is used, then the above design procedure will result in a
reduced-order controller.
V. RECOMMENDATIONS:
a) The straight application of design may be of limited use.
It is more advisable to use the proposed method delineated in this
report for developing robust contol laws since the H„/LTR design
parameters can now be properly selected to achieve approximate loop
transfer recovery in conjunction with loop shaping. The H^j/LTR design
can be considered as a subproblem of Hj/H^/LTR which is currently being
developed at WRDC/FIGCA. Choice of control design methodologies is
obviously dependent on the performance specifications of the control
system. It is believed that there are distinct advantages and
disadvantages for each design method encompassing LQG/LTR, H^/LTR and
Hs/H^/LTR among others.
101-18
b) It is generally accepted that low- order controllers are
preferred to high-order controllers, given comparable performance. The
design procedures presently being used give a compensator of the same
dimension as the given plant. The issue of controller reduction is of
paramount importance to be addressed so that the design and its H2/H^
extensions can be fully fruitful in aerospace applications. As
envisaged in this report, the significance of the compensator being
an observer-based controller cannot be overemphasized. Firstly, a
reduced-order controller design can be developed using the available
theory pertaining to reduced-order obseirvers. Secondly, doubly coprime
factorization (DCF) is a proven tool to represent observer-based
compensators, Thus, the controller reduction method based on DCF should
be fully exploited to develop reduced-order suboptlmal controllers.
The preliminary results as presented in this report pave the way to
continued research in this area. It is also of practical Importance to
evaluate the merit of other controller reduction methods in the
framework of and H^^/LTR output compensator design,
c) The mixed H2/H,^ compensator design requires the solution of
three modified Riccati equations, two of which are tightly coupled.
There are needs for effeccive numerical algorithms for solving these
equations and for understanding their qualitative properties. It has
been found in this summer reserch that under certain conditions the
H2/H„ design requires only two coupled Riccati equations. It is
suggested that these equations be used as a starting point for studying
the qualitative behavior of the three Riccati equations of general H2/H„
design.
101-19
REFERENCES
1. Doyle, J. C., K. Glover, P. P. Khargonekar and B. A. Francis,
"State-space solutions to standard Ha and control problems,"
IEEE Trans. Automat. Control vol AC -34 (1989) 831-847.
2. Fujita, M. , K. Uchida and F. Matsiimura, "Gain perturbation
tolerance in state feedback control," Int. J. Control vol 51
(1990) 315-328.
3. khargonekar, P. P. , I. R. Petersen and M. A. Rotea, "Hj^-optimal
control with state feedback," IEEE Trans. Automat. Control vol AC-
33 (1988) 786-788.
4. Maciejowski, J, M. , Multivariable Feedback Design (Addison-Wesley
Publishing Company, Wokingham, England, 1989).
5. McFarlane, D. C. and K. Glover, Robust Controller Design Using
Normalized Coorime Factor Plant Descriptions (Springer-Verlag,
Berlin, 1990).
6. Mustafa, D., "H^-characteristic value.," Proc. 28^ Conf. Decision
and Control. Tampa, FL (1989) 1483-1487.
7. Rawson, J. L. , C. S. Hsu, H. H. Yeh and S. S. Banda, "H„/LTR: A
loop shaping method for output feedback compensator design,"
submitted for publication.
8. Ridgely, D. B. and S. S. Banda, Introduction to Robust
Multivariable Control. Technical Report: AFWAL-TR-85-3102 (Flight
Dynamics Laboratory, Air Force Wright Aeronautical Laboratory.,
AFSC, USAF, 1985).
9. Stein, G. and M. .\thans, "The LQG/LTR procedure for multivariable
feedback control design," IEEE Trans. Automat. Control vol AC-32
(1987) 105-114.
101-20
1990 ,USAF-UES SU14MER FACULTY RESEARCH PROGRAM/
^GRADUATE STUDENT RESEARCH PROGRAM-
Sponsored by the
AIR FORCE OFFICE OF SCIENTIFIC RESEARCH
Conducted by the
Universal Energy Systems, Inc.
FINAL REPORT
A FEASIBILITY STUDY ON INTERFACING ASTROS WITH NAVGRAPH
Prepared by:
Academic Rank::
Department and
University:
Research Location:
USAF Researcher:
Date:
Ming-Shu Hsu, Ph.p.,
Associate Professor
Mechanical Engineering Department
University of Portland
WRDC/FIBRA
Wright-Patterson AFB
Dayton, OH 45433
V. B. Venkayya, Ph.D.
July 20, 1990
Contract No:
F49620-88-C-0053
A FEASIBILITY STUDY. ON INTERFACING ASTROS WITH NAVGR3VPH
by
Ming-Shu- Hsu
ABSTRACT
A computer program referred to as ASTROS (Automated STRuctural
Optimization System) was developed under contract by the
Flight Dynamics Laboratory at Wright -Patterson AFB, Ohio.
ASTROS employs the well-kriowh "Automated Design Synthesis”
(ADS) procedure and optimality criteria methods, in addition
to the finite element analysis, to provide an optimal design
for interdisciplinary applications. Since its first
introduction in 1987, ASTROS has received great response and
a growing user -population, and has set a revolutionary
milestone in the field of aerospace structural analysis and
design. However, the lack of a pre and post processor makes
it inconvenient in preparing the model input data and in
interpreting the results. This project investigated the
feasibility of interfacing ASTROS with NAVGRAPH which is a
general purpose geometry modeling and mesh generation computer
graphics package. Three phases of development were
recommended for the short and long term goals .
102-2
ACKNOWLEDGEMENTS
I wish to thank the Air Force Systems Command, the Air Force
Office of Scientific Research, and the Flight Dynamics
Laboratory in the Wright Research/Development Center for
sponsorship of this research. I would also like to thank
Universal Energy Systems for their concern and help in all
administrative and directional aspects of this program.
The Structural Analysis Group (FIBRA) has a wonderful group of
people to work with. I would like to thank iiT. S. Rasmussen,
L. Warner, and R. Kolonay for providing valuable information.
Special thanks go to V. Tischler for her daily help in both
technical and directional guidances. Finally, and mostly, I
would like to thank Dr. V. B. Venkayya for providing great
support and encouragement .
102-3
I. INTRODUCTION
Computer-Aided Design (CAD) systems are currently used in
virtually all sectors of engineering design practices . This
has improved the design quality and shortened the development
time. The conventional CAD system takes the finite element
model as input data and produces the analysis results. It is
the designer' s resposibility to analyze the results to make
sure that the design criteria are satisfied, otherwise a
redesign procedure will be repeated. In the field of
aerospace structural design, because of the nature of its
complexity, it is very difficult, if not impossible, to
achieve an optimal design without the aid of an automated
procedure .
To improve the design procedure, a new computer program
referred to as ASTROS (Automated STRuctural Optimization
System) has been developed under contract by the Flight
Dynamics Laboratory at Wright-Patterson Air Force Base, Ohio.
ASTROS employs the well-known "Automated Design Synthesis"
(ADS) procedure and optimality criteria methods, in addition
to the conventional finite element analysis, to provide an
optimal design for interdisciplinary applications .
Since its first introduction in 1987, ASTROS has received
great response and a growing user population, and has set a
revolutionary milestone in the field of aerospace structural
analysis and design. However, the lack of a pre and post
processor makes it inconvenient in preparing the model input
data and in interpreting the results . It is desired to
interface ASTROS with a computer graphics package in order to
enhance its pre and post processing.
1024
!
My research interests- have been in the area of integration of
autpihated design and manufacturing systems; In 1989, I
est^lished a Computer Integrated Manufacturing (CIM) program
in the School of Engineering at the University of Portland.
The objective of this program is to study interfacing a
variety of hardware and software in an interdisciplinary
environment. My background in engineering (e.g. finite
element, optimization) and computer science (e.g. computer
graphics,, data base, network), and my experience in systems
integration (e;g. developing a postprocessor for a CAM system)
contributed to my assignment for this project.
II. OBJBCTIVg OF THE RESEARCH EFFORT
Currently there is no pre and post processor directly
available to ASTROS. Indirectly, because the finite element
analysis module contained in ASTROS is mostly the same as
NASTRAN, the ASTROS user may take NASTRAN input bulk data
which is created from any NASTRAN preprocessor and with a
minor modification to create an input file for ASTROS. On the
other hand, the ASTROS user may extract results from the
ASTROS database and create an 0UTPUT2 file through a
Translator. The 0UTPUT2 file, in turn, can be used by a
NASTRAN postprocessor to generate graphical displays of
results .
There are many computer graphics packages capable of serving
as a pre and post processor for ASTROS . Among them, the
Structural Analysis Group in the Flight Dynamics Laboratory
(FIBRA) has preliminary chosen three: PATRAN, I-DEAS, and
NAVGRAPH. The interface of the first two packages with ASTROS
and the development of the Translator mentioned above are
being develoed under a contract with ASIAC.
102-5
The third package, NAVGRAPH, developed under contract by the
NAVY, is a general purpose finite element /finite difference
pre and post processing software package that combines
geometry modeling and mesh generation with computer graphics .
The objective of my assignment as a participant in the 1990
Summer Faculty Research Program (SFRP) , therefore, was to
investigate the feasibility of interfacing NAVGRAPH with
ASTROS as a pre and postprocessor.
III. EVALUATIONS
Prior to this study, many engineers in the NAVY and the AIR
FORCS had experience using NAVGRAPH for pre and post
processing of both small and large structures models. A
collection of their comments, suggestions, and the responses
from B.Y.U. (NAVGRAPH developer) are attached in the APPENDIX
for reference.
In my evaluation of NAVGRAPH, graphical capabilities, command
funtionalities, and user friendliness were evaluated from the
user' s point of view while the degree of easiness or
difficulty on the interface development was emphasized from
the programmer's view. In the following comments and
suggestions, items a, b, and c relate to the former aspect
while items d, e, f, and g relate to the latter.
a. There are three documents supplied by B.Y.U. ^ User's
Manual, Programmer's Manual, and Tutorial Manual. The
Programmer's Manual was recently revised to include Database
and Geometry sections . Both Programmer' s and Tutorial Manuals
are helpful to the user, however, the User' s Manual lacks in
proTiding an overall guidance for using the software package.
It is recommended to enhance the User' s Manual with a
102-6
Procedure Guidance which will list a sequence of steps
(commands) to be followed for pre and post processing.
b. NAVGRAPH provides the great flexibility of using Global
commands which can be selected by a key word from any other
menu. In contrast, all the other commands can only be
selected from the present menu. This causes great
inefficiency when the pointer needs to be moved from one
branch to another to select a command. Since all commands are
built in a hierarchical structure, it is suggested to
establish "hierarchical” Global commands at different levels.
c. On the X-Windows version of NAVGRAPH, commands may be
selected by the mouse, but the data still must be typed in
from keyboard. It may be more convenient for the user if the
data, either 'digital numbers or characters, could also be
input from the mouse.
d. As mentioned in Section II, ASTROS may be interfaced with
a pre and post processor via an 0UTPUT2 file. Generating the
0UTPUT2 file, therefore, is a key element in the interface
development . To understand the format used in the 0UTPUT2
file a FORTRAN program has been written to read the binary
0UTPUT2 fils and to print the data blocks in ASCII form.
e. The Reverse Formatting module of NAVGRAPH can only process
the following seven data blocks from the 0UTPUT2 file -
EXEQIN, GPDT, GE0M2, OUGVl, OEFl, OESl, AND OESIA. These
seven data blocks are used for postprocessing geometry and
F.E. models, and static analysis results. The postprocessors
for dynamic analysis, aeroelasticity, and optimization are not
available at this time.
f. NAVGRAPH consists of about 2000 routines written in the
FORTRAN language and employs a dynamic memory allocation
database developed by Sandia National Laboratory. One
102-7
difficulty the user may encounter in reading these routines is
that the variable names are ihconsistant throughout the
program. No other major obstacle has been found for interface
development .
g. In general, NAVGRAPH has the capability to perform the
postprocessing ASTROS needs in both the analysis and design
(optimization) modules. The interface development for the
analysis module may be accomplished by some modifications in
the routines, but it may require extensive changes in the
program for the design module.
IV. RgCOMMgNDATIQNS
The development of interfacing ASTROS with NAVGRAPH may be
carried out in three phases. The first two phases provide a
short term goal while the third phase is considered as a long
term project.
a. Phase I:
The objectives of this phase are two fold: i) to provide a
quick solution to the immediate need of postprocessing the two
basic types of analysis - Staics and Dynamics. ii) to use
Dynamic Analysis as a testbed to study the NAVGRAPH program
for the next phase development.
By adding a Translator between ASTROS and NAVGRAPH as shown in
Figure 1, the two databases of these two programs remain
unchanged. This approach provides a quick solution to
interface with minimun changes in both programs . The detailed
work of this phase may include:
1. To develop a "Translator” which can extract data from
the ASTROS database and generate an 0UTPUT2 file.
2. To understand NAVGRAPH thoroughly, especially in the
102-8
data base management.
3. To modify "Reverse Formatting" and "Animation"
modules, in NAVGRAPH.
b. Phase II:
The objective of this phase is to make NAVGRAPH a complete pre
and postprocessor for ASTROS by adding the capability of
processing data from the Aeroelasticity and Optimization
modules. Conceptually, this phase may look like an extension
of Phase I by only adding more capabilities to the Translator
and postprocessor. However, the complexity of modification is
much greater than Phase I. It is not clear at this point that
the 0UTPUT2 file approach should be kept or be eliminated. At
the completion of this phase, ASTROS and NAVGRAPH will still
remain as two separate programs as shown in Figures 1 and 2.
The detailed work may include:
1. To modify NAVGRAPH' a preprocessor so that it can
generate input data in ASTROS 's format.
2 . To resolve the approach to be taken for this phase
development .
3 . To expand the Translator to include output from the
Aeroelasticity and Optimization modules, if
necessary .
4 . To study the graphical needs of these two modules .
5. To understand the ASTROS data base structure.
6 . To include output from the Aeroelasticity and
Optimization modules in NAVGRAPH' s postprocessor.
c. Phase III;
The objective of this phase is to merge ASTROS and NAVGRAPH
into one single package. The data files transfer between
them, then, will be transparent to the user. The merging of
these two programs requires the elimination of one of the two
databases as shown in Figure 3. Since ASTROS employs a
superior structured data base, it is recommended to revise
NAVGRAPH to adopt ASTROS' s database,. This phase requires an
102-9
extensive effort to study the methodology of converting from
one database to another, and to develop the code. The
detailed work involved in this phase is beyond the scope of
this project and cannot be outlined here.
V. PROPOSED SCHEDULE
Oct .
1990
June
1991
Aug;
1991
June
1992
Aug.
1992
Phase I
Development
Phase II
Feasibility Study
Optimization
Aeroelasticity
Phase III
Feasibility Study
Development
RIP
SFRP
RIP
SFRP
SBIR
SBIR
—f - 1
Aug.
1991
Jan.
1992
I
I
Aug.
1994
RIP: Research Initiation Program
SFRP: Summer Faculty Research Program
SBIR: Small Business Innovation Research
102-10
NAVGRAPH
Postprocessor
Phase I
Phase II
Phase III
Figure 1 . Three phases of interface development
102-11
Figure 2 . Phase II development
Figure 3. Phase III development
102-12
APPENDIX
102-13
0*0
0 0«4
*0
a
4 OB
M >tK >«4 WOO
P 4
P P 0
•k
-k P 4
H 9 «d WPP
*0
•0
H
k
k 0 6 P
d
_ 9
4_
0
V)
Xd
• 4k a atf
4
*0 k
4
« 00
0 9V M 9 ak c
OP
d i 0
Q
4 k
HP# c Ol9k
0
p
V X
4 4H
4
H
0^.4 a
d UP k C>
3
d
d 4
0 9
0
X •-k
4 4 X 0 4
P
•k
4
k 4
0
4 0 4
P a * k £kd
a
ao
•k 4
4 «
«
HP 0
k P k 4 0 M k
0
X c
4 4
P 4
44P C
0 9.4 k 3 *4P 4
4 p
4-k
0 U 0
4 4
4
> k 4
P 0 > 0 4 a CX
••
4
4
4
Tl 4P
4 4
d
-kg k
6d 4Pa 9k
4
0 P
^d
k d
p
9 OH 4
0P444S« H
C
4 d
C
d
uie
Vd-kH
Uk k 4 P P a
0
<M 4
4
0 4
0
d
4P 0
C 3 4 U 4 k
•H
k 0
4
^•k
4 4
«-k
p
k P
9 40P0a4P
P
9 4 C‘
3
P
4
HJ3
H
4 4 0
4 ak ap > -
a
11 C 4
P
ap
tf H
9P
3
MH M 4
a‘oaa4 c
0
•k P
4
O-k
H k
0
4 0 9
HC4tl9kk40
“Oik
d
2 0-
3H
4
k 4
P 4 4 4 4 4k^
4
d 4
4
0 d
4
V
4-k 4k
d i 4 k k P
d
4 4 d
4
ss
344
^ P
•k
4 Od
4>ik ao 4 p XH
•k
ik
c
0 X
4 a
U k
•
X P 4
k Od k 4 4 4
4«4H
•k
?
P 6
•k 0
0 4
Ok k
H 0 H k 9 4d P «
4 0
H
k
-kHH
d 6
4H
CO 4
• 4 8 C 4
P 4
HH
3
d
Pd
4 d p' a>
4
'k^d
P
xd
« 4
P
4 4
d 4 ap •
4d • P 4 0 k 4 0
d
^ c 0
d
4 P
4 4 3
4
P 4
0 4 6 H
BP 4 0 in 0 d
0
4'k
4
VJS
Cd
4 k
kV 0 4
0 0 4 4
•k
d 4
0
P
d p
0 P
M-k
d 0 k 4«
4P kd 0 4 «d
fa 3 d 0 p
P
•k
P 4 k
Qdp
d P 4
d
4
p
4 9-
52
4 4
i 4
6 p
•k
P 0
ao
4 •
C 4
40
3ca4d
p 0 0 0
*0
C*d4
4
0 0
O-k
0
odP 4
BP 9 H * 4
4
4
4 d
U d 4
p Cd
*k C a T9 X p
•
4
d^M
«d
k 4kP
Skip H£M M •
4
4 4 3
3
4 0
4
4-k
0 kH
4 0 Ik 4 9Ud 4
i
9
4
4-k
m Cd
a
p Od 4 c
Pr% k 4 0 k4 W HH
0
P K U
k
P
S 4 P
:a
4H
4H 4 4k
4444 3 M 4
4
f 4*0
4
u
d9
0 0 4-0 9
k U M 0 k
aidkA H >
d
4 U
0 P
3
4 4
Si
4 PHkPHI< 4 4
4 kdkMZ4H
9ddBP 109
4
0 0 U 0 k
P 0
k C
•
d PP k p 4
k
K-k 6
p^
C 4 >•
4
9 - -
k <k
OP
k • 4‘OH
e P If 4 4 T9 9
•k c d
o* d
4 k
0 0 4
P-k
Htf 4 4
4 9PPH a c c
0
POO
0 4
M 0
H 4 k
4
6 k P P
B k 0 0 ak(A 4 0
P
a«4^ 4^ d
ii ^
4 u
4 -
4 4 4 4 4
4 k a BOH g
0 PP 4 P-k
HP
5 4>
M 4
HP
4d 0 E
H Pd EH 4 QP
4
aad an
H
(Ad 4
H«k
d 0
pa kk
HCP 4X* C4
M
P 0 0<k 0
•k
PK
H k
-k C
0 4 'OP
4 4 0 4 aokd
•w4
4
Ik k
tf 3
9
4
09 CH
k ack a Q 9 p
4 0 4 OP koS a
H
H 4 4
4 4
HP
b
V •
4<M
« 4Hk 9
d d k cd
9
C 4
H
OH
C6 9
> PP k 0 13 d
^«H«k 4^ P
Q 4
-k 4 •
9 4
a
k 5 4t>
X >»3dH
0 419 ka 4 »
«k
tMW<kdH 0
3P
4 MP «
0
4 4 0 E Ok 9
9
P d
4
9 4-k 0
3 0
p<^
H p 9
4x 0 4 Q Pd 0
9
k
- > 0
p
H
4 4 P 5
dk p 4 •a p k
3
H 9
0 4 -H
H
gPk 3
pH d 4 B WTf d
4d u
a<M
4 k 4 4
H
4 C
0 9 4
9 4 3 4 P
H
*0
k 0
4
S
UIH »>**d
T9 6 9 4 k N P
0 .
i
g
0 0 4 0
Hk BH D»Z 94a
• UP
«H 4 C
♦ k
• B
« 4 4Hk
• 9 > o5 a 4
«
069
H d k 0
r4 0
H 5
«dH 4d
iA0404k5a4P
0
HH 0
H 0 0 0
H c
H 0
H 4X19 3
H 3dPd ap 0 E4
rSs'JS
:^ar
lO M W
„ S.fi
<4 « Z M
■h^-i
n> wtJ
Cit **4 CH
OfH 41 ) 4fH
ai: 4
TJ e >io
V 0 u u
« H e 0 4 0
4 41 0 3 MiJ
A M
• e « 4
£4004
C*f U*> 4 M
4 9 C«4
4 » 0 N C
A 4 UX SW
a4 4
4H 4 4 4
>H4
4 4A ^ •
3 S# d
4H 4 0
NjC^WH 4«rl
Kl 44i 4 4 H
P -*4 4 9H N
C^3 Hfu^j
r4 U U
P c^'d 4 N
d C 4 ^ -H
0-HA >« •T;
p ej< ^ 4'd
b Ne 4 k
4
: d 4 4^
O P*k p
I d fi k
i-( ^
a ^
* * I
*j *1 ^
fi ’b
5 ii
>«4 0 4 4H C
4 O
•k 4 P 4 P ‘k
kd p d 4 P
“ ■* - Pd 4
•n 4 p p 0
H H
4d au'k -H
B 9 W 4 0 444
H a
4 9
9 U Cfk
u d (A
0 0 4 6 k
•H 6
4 k 4 4 ik 4
> 9
ddHd OdH
• 4 P
HP 4 P UP 0
Hd-H
4
'd
p
4 M 0*0 M 4
d 4 P C>k p
9 >k 4 U » U
p w k a>k
3 4 a 4 9 9
•o d If
P P 9
•k 4 O 9 U
• W<U g
«Qr-l 0 0
MH 4 UP
C V 0
o V e OA
9 P P
k O 9 _
a 0 4 4S
o 4 3d pa
o» p aS
CM e3
4 4 POO
9 k 4 M >
0 • a<
•k k 4 4 X
k 0*0 k 4
dM C k 4
p k O d
S4 4*M O
k d 3 4 W
0 4 0 4 OP
'w Hd c a
P'k o
p p
U UO O'k 9
p 4 d d 4d
•k^ 4<k 3 p
e 4 *0
•k 4 kT9 0 *P
4
4 4
MM
§3
4 •
P k
p p
4
'M O
0
9
d
•0 P
4
4«4
P 0 •
4 H
0 4 4
-g|
M'H 6
5 UO
k 4 I
k kr^
4-k
V 0
4 d
4 UP
W<k*k
0 4
4 d
^ P
4
4
C
0-k
P 'M •
4 P
4T) d
U 4
<k M
C 0 4
P k
4 a
d 4
4 P
«k C 4
9 4
o d
5 k-k
0^
p
•k 4 O
*0 d
•k “k
«k 4
4 09
*0
C p
4 «4
f« 9
kr»
O ® ^
U p 0
U p
a 9
w >0
u u
3 HJ 4
.0|
H
H 4 4 P 0
p
1 4 0
H
0 a
0 4 9
w
4 •
9P ak-0
PEC
9 4
P 4
d 9 0 4 P
4k
H 4
0 P
k 0 U 0
4 k 4
a 4
P V
kH p d OH
X 4 9
Ud 4
H
0 A 4 0 a
k ak
0 P 4
P 4
p 4 9-H a 4
H C
X
0-0
d -i-kp 0
4 0
4 4
c
4 4 a
d u
d 4
•0
A 0 0 0
TJ P
P PH
g 4 AT9
^ c
P W P
H 4
Ok 4
6 4
C 4 4
9 4 w
MH k
4 0 T) d A
0 p 4
a a ak
H •>
A p • p 9 4 •
3 4H
EHk
4
3 U C k P
0
OH U
w
4 0 4kH 9
k d
0 4 4
4
u a> k aa
0 0
• 0 4 a
• k
• .*4 a 4 p a c
• d 9
^ 4 k 4
in 4
\0 C 0 d 4 4 k
r-k g
4 4
k k C
0 0 0
lu a-k
k P
0 4
U P
• C 0
C0*k k
102-14
-li
m 0 't
>.• S «
55
ini
®i!i
IW
a*«j
hh
3 9 • ■
O CAW
A f a u
c'-i
“S83
il li H
?a*'*
JJSw
• **u
S 2
: 5!
• tt« 9»
e
if<;0 •
• 3
ijiS."
8|-H
AO'I’j''
**■5 3
o • '
i e#4
. a'",h
• 9<iAa
5s ml siii^
I. h' lli; 1 1 I;
1 1 ij Ij I
lii:-
m
’Hi
If
.s .
m
£8 i
m ^ « « ■ 4i
U
Sl«^’
a 8 « • (
’:ui
«>3<^
A'V >«•
.5:8.5
AM 9 48
8
8^85
ftlto
8|?h
ji|a
S»|S
111]
.Sja
c • Sa
oS«S
e ^
Ak44#4
« • 3 9
sS5l
N- 41 • -
5858
341 CM
• •9 0
c e«4i
4> 0*0
« 0
U
8^58-
Is?"
fiS?
M
^..5’* 81. 8 8
slaa si :;>
l|p. i jl
lii!' '* ■'
Igisl
182158
^1 !p
sa I? 8 •
Si 585
■B « • ” !
.C-.b',4. !
•?5'S, 5
&5SaN ft
o^ • 8 •
S' - w
.15- -5' S
5-. 8 la S;
5iS|8 8
i^sh f
•IMS *
g
'§^
5% 5 1 8
^SS85 i
^th I
I
8fa^3 9
35** * • «®
ml
Is
882o» a.
*5S a .
5858 13
«ai .8 <n3
»7o *8
Ull8
t
UU M
sc H
’^i >
•, %a Q
M9 A
rs g
lift I
3 T-»
O - ^
MO 3
Ah U1
8|5:
5*1^
^ 0 0
41 H
5j**'8*
|a'*S
■ ■ 8
■
aog
a.x^
0 H ij
0*9iii j
O 0 H O
*54
cLSS:
“gSc
A?* • •
5|«g
ggsa
^ V W
F>'S - •
M « O 0 0
o 3 aj: T8
n A 04^ O
0 U
>«M 0
U 41 0
• 0 O C£
H > 4J*H 4*
4»0A «
H9 0 0
S do
0 0
0 00 M
0>0 A
.**858
•H s s
S . u
a ’a
a cK a
9 • •
Sa ■ '^
•i 8 w
.J) M 0
I ill
5'-^ AC
•t 0 8: c
852?
3- Mr1
w affo
Vq C
-•85
0 A
MO 0 M
01^^ O
0 9^
0 0 0
»■« 3 4f O
■ • 3 3
»• 88!
3^* *“8
a. «aa
s fMa
O S £ 6
!** 8s2*
: a ga ft
a ■ " o
> n u .a a
: . Aa o>
J* 5*'®a
r5 g?55
• f c •
§l»s
M 85
S''«5
ilSfS
„|iA
”s8"
S;58
0 0
a.a,c
s«l|s
isPS
:s:si
Sss|;
p^iJ
|5i*1
5 58 S3
^.588
% A O
g 0 »0 0
^4; c w ,
9 U'Ti u M
048 3 O
mao 0^
48 0 0
4i O O
8 S S
: s :
• =■ =•
•r4 0 0
M M
0 3 3
•H O O
§ £ &
102-15
9t-Z0T
U • «
• ^
> 4 tt
O.U 9
•-t 4 if P
a 4 a
nb B C
V 0 b
B e bi4
0 B a
9 P 9 9
VC C b
•H u i;
0 ct* c
P V 4 4
B.B *'0
«'b • •
4 0 b-
0
u
P 0
e
0 b
b ft B
a 9 0 4
ft
0
b
■4
H *4
B
ft
B
H # B
V b
« 1
■ 4 n 4
4
0
0
0
©
X
*4 b
£
3
B
3 ft ft
B 8
9 ft
0 b hi ft 9
0
ft
9
ft
ft
^ P
b
X
X
0 ft B
ft B 0
4 b
b 9U C
H
a
0
C
9 B
<
4 P
G
ft
ft 3
ft U ft
i) 4
a 4 < 4 B
ft
ft
4 *
0 0
X
^ 6
V
4
B
« 0 C
b 9
4 4
4 w 0 X
0
b 0
A 5
u
1 I
0
ft
ft
ft
V 9 0
ft 4
c X a ft
b
0
0
6 4
>
U i
«
B 0 ft
b 4
X 0 9
e
t
4
4 P
*
‘ i
ft
4
B
4 4
«
e ^ 4
• ^ 8
9 0
ft ft 0 0 0
0
0 r.
P
U
» ^
• u
X
« V
•H
„ B
U
4 4 4
9 0
9 4
C 9
4 0
<M
% ft
8 P
• ft
P
8 4
c
B C b
b d 0
B
« 0 tJ 4
0
4 -4
t* *
4 U
• »•
V 41
b
P
b i
4
i 9 B
4 C
4 «4 ft 4 •
b
M ft
"p 5
S 4
} ..
b
0
:b M
4 4
ft
0 b ft
0 B 4
9 9
B 9 hi 4 ft 4
0
H 4
0 B
0 a
• J
ft ,
ft
b H
4
’U « 4
ft i
9 4
•4 0 ij B a ft
0 ft
■ >4
ft 4
, «■•
0 ft
e
V ft
9
» 0
b
. B i
3 3 8
8 ^
ft X hi c i e
0
Q 8
. b
ft
*»
0
2 ?
M
0
B ^
0 0 4
•4 B
V 4 0 "4 0 3
0
3 0
B <4
ft »
> >w
■ •
4
0 w
»
> k
W 8
A U
ft 4
ft ^ H u 0
X ft
0 tb
• •».
0 9
b
1
4
p i
B
«-* B ft
B B ft
9
e 4 B a a
>,
4 0
e
e <
(
© b
4
ft 4
8
b
ft
H 4 B
ft 4
t H
B B X 0
ft
X 0
0 B
c 0
u
e
H •
4 «
ft
X 4 4
0 ft a
X 4
4 ft B 0 b
ft
b
ft
0 •
1 . 1
* H
p
-4 0
4
U ft
hi ft
c
4 >
Oft ft X 0
4
0 0
3 ft
0
1 J
e B
1
4 0
H 4
X
A U
4 « e
4
H «8 • 4 ft N
X
ft X
0 4
>.
4 B
4
P 3
B
b
0
0 4 ft
9 M 0
b 0
U ft B -ft 0
0
9 0
X 3
b X
• B
X
V 4
b
4 B
ft
X ftH
8 4 4
ft 4
ft e ^ 0 0 C
ft
0 3
B 0
4 0
P A
N
b
4 *
1 ft
4 *
X 0 ft
0 b ft
• b
B 4 a 0 ft 4
9
X e
ft
0 4
A
U U
4 b
• Z
0 b
>.
t b C
U ft 4
b
4 SBC
4
0 >
e 0
l-U
0
ft
a 0
b ft
b
0 b
0 ft 4 •
4 •
0 4
0 0 4 4 ft 4
0
>•
C P
3
X *
b
'44
0 ft
3 b
ft 4
< b B
hi B b
ft 1
4 ft X 4 0
ft
0 ft
4 0
0 u
n
4
b
X
'4 Q
4
•H ^ b b
X 0
4
9 b 0 ft C
C 0
S C
X 0
•
C
4 Z
0 0
ft 4
X 4 3 B
0 ft %4
ft b
ft 444
>•
4 >
1 -4
«H
X ^4
P
H ft
V >
ft 4
ft B
hi 4 U ft
ft 0 4
ft 4
9 0 b ft H H
ft
B 4
0
9
B
a
p ft
ft
4 ft
•4 0
H b 8
0 4 C
p a
c 4 0 4 a
■ ft
0 0
C 0
h 4
b
X
ftz
3 8
H 9
< B 0 3
hi ft 4
•4
*4 0 hi ft 4 e
0 9
c
ft ft
9 A
M
90
b
1 0
P C
E ft e e
X 4 b
4 0
ft 4 9 0
i4
0 0
hi 4
>«X
9 0
9
c >
«4 9
ft
ft
4 4
B ft
ft ft
4 9 * 0 4 ft
X
9
H
ft 4
*•
0 *4
b
b ft
p C
z
4 B
H 8 B B
9 b
4
ft 4 • X ft
4
H 4
hi 4
aft
C
Ct s
ft
X*
P P
ft 8
ftv
0 B B
W 0 9
aft
0*0 0
a
0 8
*3 B
a ft
a
0
ft V
< 4
4
ft 0 0
-* P A
4 U
0 b e ft 0 3
4
•3 0
hi 4
4 4
0
p4 0
4 3
X b
0 4
•390ft
> 9 P
U 4
4 4 aft b
0
a 0
0 ft
>
z
p b
^ 0
0 9
4
ft b
e 9
a
e 0 0 • 9 ft
1
ft
e 4
ft
0
B B
> 0
8 i
B ft B
999
4 9
0 b 0 0 0 6
4
0 8
B ft
B
(K
♦« ««
•*
V 0
ft X
< b
f 8
ft a ft 0
A U P
A 9
X 0 0 b X 0
X
X 0
X 0
X B
0
B
s
0
n
0
B 0
ft B
S ft
H e
H 0 ft V
ft ft >
ft w
9 I u a u 0
H
H c
H B
9 X
hi
■ p
9
8 0
1
«
«
•
f
E
’*» ft
0
ft 4
«4
M
<n
9
0
ft
•
9
0
ft
. 4
omZ
OT>
••mh*
o I 0
«4J ^ 4
• • M C
t) 4 0 9
> b
BOB
*-4 C
ft 5f 0 '
• ® i
*4 4) B
^ B B-J
B V ft
*0 B B B
c y b
B 0 B *0
b C
B ft B b
C i
••4 B 'M
M b e
B ft B ^
V B B •
oi^V
V «> B 8 •-*
B
S e X B B
0 *«• JB
«B P 9
b B *0
V B ^
9X 0 0 >»
OP 9 b
•H 4i f
b * * B ^
0 #4 • 43 B
0> B U 9 'H
C B
«H 0 O **
8 B B B
0 ♦ *9 B
B B B 0
•a 6 8 P
B 44 0
^ 0 8 44 B ^
a B **4
B V 9*0 44 A
• 8 B
^ B 0 ft
0 43 B •»« >* «
aft 4
ft b 4
9
C 9
9 _p
0 «i 4 B
*' ■:>
X •: B
ft
ft
4
9
B B U
0X4
ft • X
ft
4 b
0 ft ft
0 4***
9
3 0
b ft 9
B b > 4
0 80
«M ft ft
U
•ft
< a
1 0 P
9
P
0 ft ft b
ri 4 •• ft
B 0 0
p
A 0
B 9 b
X 4 X
•ft
M a
Q 9 !
0 e
HX 4
Q
In 0
B > ft
«»
B
B
A UP
BOOB
j: B44 B
P 8
B « B
C b U
h4 b B B
0 M ^
d 9ft
0^ 0 B
V B 4»'0
B b
« 44 B ^ B
b 0 ft B V
0 B e
44 B e B B
c 41 M b B
•H B P 9
d B^
44 >*4 • B
0 U d B B
B
ftM.-o-j
B B -b 0 d
p p *0
B B B B
> • *4 V
B*0it 0
>i^ B B B
B ft >
^ 8 4 B B
ft B«H 41 B
B M ft
•H B B 0 ^
*0 •*« 4^ B
b*0
0 0 * B
♦> ft B B V
9P 9
B tj 4*0
p^ • 0 *0
A B C ^ B
B B B U
. B B 4' b
9 AA^ 0
X 9 P
P B
0 B 44 B b
^4*0 0 b 0
j
K 44 pi B ^
VI S B P 4> .
*B 9 B®- ^
55-**t
> B 44 BM d
p 9 P ^ B J
9 PV'**Q
5 B b B ^ b
8489 80 0 4
UP X _ 8
ft *71 ft
«mPb dft b
B B 0 < B S B
B •*« < ft ft -
U V B H « ft ft
d d 0 i« o b
B B B ft Q > B
V BftSSft b
« B«> ^0« ^
C B 8*^8
I 9 E 0
B 0 b M ^
ft* b K d J*
d 4^ 44 d U '*4 U
B M • a ^ •
B B UU B p
B Ob
C • B B B*«« ft
8-8 8 ft^ ft
b B 0 B B B
o 0 4» b <-•
• V B V b B B
U U^O B B e-N
b B 4S P ft
0 b B V 9 b V
h^P ^CP
8 B B •H ^ *
b 0 b d B b
•H U B B *4 8 B
< b X B b >
8 B B V B B
B 0 C 8 > S
• 4: b B 8 0 B 0
I 9-*« U b S
d • 8
0 B b
b B B
PUP
ft B ^
B > B
h I 0
0 p
P uA
*4 u ft *
Z 4*0
44 B b B
0 0*0
^ B
ft 0 8 B
C OB
B d U
8 0'*4 B
^ M h
p p^ 9
9 0 0
OB B
b0 8
*0 *4 d B
B B B ft
*Sft S "
(kp 0 ^
p 9
9 CP 9
9 ‘*4 9 9
ft Ifft
9 8V
O dft
M •«4 b M
HISS'S
S**u
ft 8 A 0
B -M H <M
B B U B
d 0 0 >
BO b B
0 *S <8 ^
B ft (8
I V 43
w • *0 B _
B b C 0*0
b ft B ft
> I
0 8
9 P
8 8
*4 B 9’2
Bft<«4 B i
b 0 *4 0
;:8^2|
"ullUS
• ° ••o ***
tuS » >>i
K K S • ><
OS 0 it 8 •
> > V A 0 ■
«C <B *4
B B B ^ V B
0 B B V
A 9 OP 9 e
p Cppp 9
•«4 *4 b d 8 f
s B f B -H I
9 0 V V 0
*0 9 ft 0
B b ft© _
B B • P
8ft • •ri S J
3^3»£§S
•^•5*522
? 82 S
b b b ft 9 >
!*SSC52
C O 0
8 Oft> ^ B
94< < • B
B 9B w« 9
ft **4 0 O
a b b 90 M 0
iH *4 4: B O ill V
ft V '*4 o( V
^ BHU B
U b B 9 0 Q B
Q 9 B ft « 9
B 0 *«« 0 3d
B V V b ^ *4
CP 0 ft V
%‘H p4 d «M 43 d
^ B 0
B B43 > b*0 0
b b B -H 0 B ^
• B ft 9 B ftft
C B B B 0 a
B C 0 9 B *2 r*
0»*«4 C 0 B ft
9 B **4 0 >
d C V B b B B
M B •H 43 0»V B
•• «• B s
•
V Oft •*
B b 9 0
Qft0 H
n ^
il 1 1 P
i I ?
1
is
Hii!
lllllilHIi
|.|
l&i sJsssgs
J
I .
^ « « ♦ w
U I
II
I f
If a
-i 8
I II
I M
• is« •
Ifil III
. I
f M f»
10
^ <0 K
102-17
1990 USAF-UES SUNWER FACULTY RESEARCH PROGRAM ' ' .
sponsored by the
AIR FORCE OFFICE OF SCIENTIFIC RESEARCH
conducted by
universal energy systems, INC
FINAL REPORT
theoretical modeling of the perforation of laminated plates
BY RIGID projectiles
Prepared by: David Hui
Academic Rank: Associate Professor
Department: Mechanical Engineering
University: , University of New Orleans
Research Location: WRDC/FIVST
Flight. Dynamics Laboratory
Wright Patterson Air Force Base, Ohio 45433
US Air Force Researchers : Greg Czamecki , Effort Focal Point
Date: 9/30/1990
Work Unite No. 24020261
Contract No.
F49620-88-,C-0053
THEORETICAL MODELING OF THE PERFORATION OF LAMINATED PLATES BY
RIGID PROJECTILES by David Hui
1. INTRODUCTION’
The present work is concerned with the penetration and perforation
of laminated plates made of composite materials. The velocity of the
projectile of interest generally lies within the ordnance range, that is,
from 500 m/sec to 1000 m/sec. In additon to the many complicated phenomena
which are associated with the perforation of isotropic -homogeneous metallic
plates, one would need to consider additional vastly different in laminated
targets and they can be grouped into three stages of failure: matrix
cracking, delamination and fibe-** breakage. None of these three stages
of perforation of composite materials is understood since one has to incorporate
the complex interaction between fiber and matrix and the wave propagation
and high strain-rate properties of composites are not well understood.
Further, due to the hugh amount of structural parameters (number of layers,
lamination sequence, plate thickness, size and shape of the projectile
and size and boundary condition of the plates) and material parameters of
the composite plate, it is expensive and practically impossible to provide
e.xtensive experimental tests involving all these parameters. As of today,
exis-ting.'-sitheories for the perforation of composite plates are empirical
or semi-empirical in nature; thus, the theories are not predictive and
would require some measurements (for e.xample, the weight of the plug, etc.)
The aim of the present work is to develope a predictive theory which is based
on the fundamental principles of mechanics- The theory is necessarily
a preliminary one due to the enormous comple.xity of the perforation problem.
103-2
The theory being developed is based on the energy balance method.
In order to assess the validity of the applicability of the "metallic”
theories to composite material plates, it, is desirable to stimraarize
the existing metallic theories which are based on the energy concept.
Of course, some of the observations and conclusions from the metallic
theories remain valid for composite plates. For example, the type of
perforation can be classified as essentially plugging or essentially
petalling, or a combination of both types; sharp-nose projectile within
the ordnance velocity range tend to favor petalling while blunt-nose projectile
or thick plates favor plugging. The remaining part of this section
is devoted to a brief survey of the metallic theories using energy method.
Starting from the classical work by Taylor in 1948 on the theory
of the enlargement of a circular hole in a thin plastic sheet, Thomson (,19.55)
was the first who presented a quasi-dynamical theory for the petalling
perforation of plates for rigid projectiles with conical or ogival head,
using the energy balance method. Thomson's theory contained an algebraic
error and it was later corrected by Sodha and Jain (1958) who also
presented the residual velocity expression. Further, Johnson et al. (1973)
examined the same problem but included the dynamic effects. Using the
results of the bending of triangular cantilever beam under a tip concentrated
force (Parkes 1955, 1958 and Johnson 1972), Landkof and Goldsmith Q985).
found that the initial stage involve the formation of the triangular
cantelever beams due to crack initiation, followed by plastic hinge motion
and petal bending and they presented the residual velocity. Finally,
Woodward (1978) extended Thomson's work to include the work done in bending
and he was able to explain the transition from dishing to ductile hole
formation as plate thickness increases.
103-3
Recht and Ipson (1963) were the first to present a theory of
plugging for metallic plates using the energy concepts. The energy
balance equation is, assuming a rigid projectile,
(l/2)mpVp^ = Ejjpj + E* + (1/2) (mp+m*) (v^^ Cl)
where nip and m* are mass of the projectile and plug respectively, Vp and
are velocity of the projectile just before impact and residual velocity
respectively, is the energy required for deformation and heating
and E* is the energy required to shear the plug out. A mathematical
model for the plugging perforation of plates was presented by Awerbuch
and Bodner (1974) who considered the three stages (i) compressive
stage where the projectile acquire added mass from the plate with no
shearing of the plug (ii) both compression and shearing of plug occur
Ciii) plug continue to shear as a rigid body without compression of plug.
The three-stage theory by Awerbuch and Bodner was later refined to five
stages by Ravid and Bodner (1983)
(1) dynamic plastic penetration
(2) bulge formation
(3) bulge advancement
(4) plug formation and exit
(5) projectile exit
Further insights into the plugging process was presented by Woodward (1987)
who considered dishing, stretching and plugging deformations where
no post-perforation deformation measurements are required (predictive theory) .
The application of the above five-stage model to ballistic penetration
to ceramic armorer ceramic-metallic armor was presented by (Ravid,
Bodner and Holcman 1987, 1989).
1034
2. PERFORATION OF LAMINATED COMPOSITE MATERIAL PLATES
As a- first approximation, a lamianted composite material plate can be
considered as stacking of the same ntunber of layer of metallic plate
where each layer of metallic plate is of equal areal density as each
laminated layer. These metallic layers are ’’glue" together such that
the shearing between adjacent layers is the same as the interlaminar
shear stress of the laminated plate. Such modeling, especially with
the inclusion of the interlaminat shear stress , has not been considered
in the literature.
Marom and Bodner C.1979) were the first to examine the plugging
perforation of raultirlayered beams which are separated or .in contact.
In the case of contacted beams, adjacent layers are free to slide (no glue).
Marom and Bodner did not consider glued beams. Note that it is not
obvious whether the the inclusion of glue would be advantageous (in a sense
of defeating the projectile of reducing the residual velocity) over
the no-glue case. The advantage of glue is to increase interlaminar
shear resistance and the disadvantage is to reduce overall bending
of the beam (note that overall bending may be an important energy absorbing
mechanism) . More discussions on this topic for unbonded layered plates were
presented by Radin and Goldsmith (1988), using essentially the same
equations as Maron and Bcdner, The advantage of "bonded" versus "unbonded"
plates is subject to some controversies since bonded plates has shear
resistance between layers (advantage) but provide little overall deformation
(disadvantage) .
103-5
The energy balance equation for the perforation of bonded layered
plate is, assuming a rigid blunt -nose projectile,
(l/2)mpVp2 = E.^ Cl/2)(mp+m*)CVr)^ (2),
The above equation is identical to eqn. 1 .vith the inclusino of the
energy for interlaminar shear between adjacent layers. Note that vanishes
in the case of unbonded layered plates or when the total thickness of
bonded layered plates is small (so that thin plate theory applies) .
In classical lamination theory, no account is taken of the interlaminar
(perpendicular to the plane of the plate) stresses such as the interlaminar
normal stress and the interlaminar shear stresses and Rather,
only the in-plane stresses o' , o' and are considered (Jones 1975) .
A / xy
Thus, classical lamination theory would not be able to predict some of
the stresses which cause failure of a composite material. Jones (1975)
showed that the existence of interlaminar stresses implies that the
laminated plate can delamiante near free edges. . .Such free edge may be
a hole ganerated by the projectile in the perforation process. Note
that the interlaminar stresses died out rapidly away from the free edge
and these stresses are significant only within one laminate thickness
from the free edge (Pipes and Pagano 1970, Pagano and ^ioes 1971 and
Pipes and Daniel 1971).' Finite element solutions for these interlaminar
stresses as a function of the stacking sequence was presented by HalpinCl984),
using three-dimensional elasticity equilibrium equations.
103-6
Assuming the worst case^ when all the layers are "delaminated" from
each other, that is, adjacent "metallic" layers are free to slide with
respect to each other, one can employ the energy balance equation to
obtain the ballistic limit velocity. Setting the residual velocity of
the rear layer Cthat is, the n^^ layer where n is the number of layers)
to zero, the work needed to penetrate this layer is
Note that the is computed for each layer such that the initial velocity
for the i^^ layer is the residual velocity of the i-1^^ layer. Thus,
one can start from the rear layer and compute the initial velocity for
the n-1^^ layer, then the n-2^^ layer so that eventually, .the initial
velocity for the first layer is obtained (this velocity is precisely
the ballistic limit for the entire laminate if one starts with v =0
r
for the rear plate). it is -postulated that,
V (i^^ layer) < v (i-*-!^*' layer)
Bl
since the size of the hole tends to increase due to the added mass effect
on the projectile. Since the layers are in contact with each other, the
energy dissipated for the i^^ layer in the plugging process is due to
the shearing of the plug as well as due to the compression of the
remaining layers i^^, i+1^^, . . . ,n-l , n^^ layer:
or (TR^) (t.+ t.
u^t 1 1+1
,+ t )
where is the ultimate stress and R is the average hole size.
103-7
Following Marom and Bodner (1979), the energy transmitted, to the
i^" layer in the shearing of the plug is the same as that for an isolated
layer so that, for each layer,
*
E
s
(1/2)
where nip is the added mass of the projectile. Note that the balli‘'tic
limit of the i^^ layer is related to the ballistic limit of a similar
isolated layer by,
YblCI^^ layer) =
2 _ *
r- , ^ j -th , s O'.. 14. (tOCiI )h(ni„ + m )
V (isolated i * layer) + ult ^ p
oL — 7 "
V
103-8
5, PARTITION OF ENERGY DISSIPATION IN C0^1P0SITE PLATES DUE TO PROJECTILE IMPACT
AND DESIGN OF EXPERIMENTS
2
The toral energy just before mnact is (I/l) m_v
r p
The excess energy is (l/l^ ~ ~
This energy is dissipated in the following form
(i) Hertaian Contact and oompreosion induced added mass Ejj (fusing MTs)
(ii)' Plate bending E ( measure using MTS and strain gauge to determine
the derlection)
(iii) . Delaraination, including matrix cracking and fiber breakage E
Rising Kopxinson Bar and then C-Scah^
(iv) . Shear Plugout (using MTS and Hopxinson Bar)
measure weight of plate before and after shooting
(v) Petalling due to excessive circumferential stress
and stretching due to inelastic elongation Ep
(^impact of piate with a pre-drilled t±h« hole)
(vi) Vibration, possibly finite amplitude vibration with damping E^
calculate oased on material and structural oroperties, witn
special emphasis on time to tirst peak and logarithmic decrement
fromthe aisplacement vs time curve, one can estimate tne energy
asorption.
c
“X
E
H
=s
E„ + E,
V
us'ng Poly-vinylidene fluoride PVuF Pie<opoiymer sensor, one can
imbed tnese gauges in oetween layers ana suoject tne composite material
sample unaar impact using the Hopxinson Bar apparatus
103-9
(i) Hertia'". Contact
Force = const, (relative approach) valid for quasi-static impact
The energy dissipation begin with the stress- concentration at the
contact surface, followed by compression induced added mass
This Hertian Contact phase will end with the formation of the
first crack starting from the contact surface.
The above power lav/ coefficient should be determined for medium or
high speed impact.
Tne aoove formula assume that tne contact auration between the impactor
and tne target is very long in comparison witn theirnaturai period .of
vibration. Thus, it should be modified for very hign speed impact
of piates so that the rorce is proportional to a different power.
The constantant of proportionality is a function of the geometric
and material properties or both tne projectile and the laminated plate.
103-10
(ii) - -plate aenaing
Circular plate unaer concentrate load at tne center
close form solution exists, assuming clamped ooundarv
conditions, one can find the deflection and' hence, the maximum
stress located at tne back surface of the line of impact.
Equating this stress to tne failure stress of tne compostie material,
which is usually slightly above the yield stress. One can then
determine tne impact force using Gres-aciuk metnod for low velocity
impact. Csee 'Shivakumar et al. 1985a, b)
(iii) uelamination
.one can use very thin aluminum layers bondea togetner and test
the oonaed plate in a hopkinson bar setup.
The initial wave is compressive until it reaches tne back surtace
ana this wave is reflected back as tensile wave. The tensile wave is
responsiole tor the aelamination separation or the lavers . such
waves can be detected using imbedded sensors. Tne sensors will enable one to
obtain the stresses versus time and compute the enerov aissioation
between adjacent layers.
. shaar Pxuoout
■i'ha iM^lt IS lubjsctsd to a fiat nosa projactila impacting tha oiatt
rasting on a hat-snapad rigia aupoort. Tha affacts of shaar piugout
and patalling can oa uneouplad by tha intreoucing spaeiailv aasignad
axparimantt. For axanpla, tha indantation or a fiat ounch, laa
abova diagram using MTS for static taat and honkinson bar for
dynamic taat will anabia ona to datamina tha anargy dissipation
oua to snaar nlugout.
(v) Patalling
As a ruia ot tha thumb • thin piata tands to oatai ano thick plata
tands to plug. Itost piatas iaotropie-haa»gana6us or laminatad-tibar - rainforcaa *)
woula fail undar projactiia impact in a combination of olugging and natailing.
Patalling is dua to tha anlarganant ot tha cyiinarical suxtaca due to tha
addad mass, such tnat tha noop strass axcaaa a critical vaiua. We can
aatarmina the enargy aua to patalling by drilling a tinv hoie and snoot
the projactiia at this hola using tna MTS machine tor the static case
ana hopkinson bar in the oynamic case. Mote that tne honkinson bar anoaratus
would enaole one to ''aim" at the tiny hoie oraciselv.
o
tiny ore-drilled hoie
103-12
(vl) Finite Amplitude Vioration
The plate wilx vinrate under the initial condition or iero displacement
and a prescribed initial velocity. Tne structural parameters will oe tne
site of the piate, tne total thickness of the plate and tne boundary
conditions. The material parameters will be Young's modulus and Poisson's
ratio for the isotropic-homogeneous plate and there will oe more such
elastic constants for a laminated piate. Furtner, one would need to estimate
tne VISCOUS camping (force is proportional to tne velocity^ coefficient
and the material damping (^ysteresis damping) coefficient using the
complex modulus loss factor. Based' on tne displacement versus time curve
of the vicration, taking into account finite deflection terms, one, can
determine tne energy dissipation. One need to determine tne energy/ dissipation
in the first few cycles since the plate tends to damped out quickly .
Tne first or second peak is likeiy to cause most of the damage since
they usually correspond to r'^latively large deflections.
D
time
103-13
4. CONCLUSIONS
The present work deals with a preliminary attempt to formulate
a theoretical modeling of the perforation of laminated plates by rigid
projectile at ordnance velocity. Since the subject is an extremely
complicated one, the present work would focus only on energy dissipation
aspect on which the theoretical and the experimental programs are based upon.
Although there are still many unresolve issues, it appears that the
important issues are spelled out and they will be the subject of intense
investigation in the coming years. Further refinement of the theroy is
needed as the experimental data becomes available . Extensive tests
using the Hopkinson bar apparatus, charpy impact tests and the static
or drop tests for impact on laminated composite materials, plates are planned.
5. Acknowledgements
The author would like to thank my US Air Force colleague, Greg
Czamecki for his technical assistance of this- program and his help in
the experiments performed by my two graduate students, John Lair and
Magna Altamirano. Special appreciations are due to Pat Petit for her
support in the experiments, James Hodges (chief of Survivability Enhancement
Branch), John Sparks (chief of Technology Group) and many other co-workers
for their understanding, encouragements and assistance. It is my pleasure
to interact and share research results with Dr. Arnold Mayers (WRDC/FIV)
and Dr. Piyush Dutta (US Army Cold Regions Research and Engineering Laboratory) .
103-14
6. references
Awerbuch, J. and Bodner, S.R'. (1974), ''Analysis of the Mechanics of
Perforation of Projectiles in Metallic Plates", Int. J. of Solids
and Structures, Vol. 10, pp. .671-684.
Halpin, J.C. (1984),, "Primer on Composite Materials: Analysis", revised
version, Technomic Publishing Co., Inc,
Johnson, W. (1972) , "Impact Strength of Materials", Edward Arnold
Publishers Limited.
Johnson, W., Chitkara, N.R., Ibrahim, A.H. and Dasgupta, A.K. (1973),
"Hole Flanging and Punching of Circular Plates with Conically
Headed Cylindrical Punches", J. of Strain Analysis, Vol. 8, No. 3,
pp. 228-241.
Jones, R.M., (1975), "Mechanics of Composite Materials", Scripts Book Co.
and McGraw-Hill.
Landkof, B. and Goldsmith, N. (1985), "Petal ling of Thin, Metallic Plates
during Penetration by Cylindro-Conical Projectiles", Int. J. of
Solids and Structures, Vol. 21, No. 3, pp. 245-266.
Marom, I. and Bodner, S.R. (1979), "Projectile Perforation of Multi-
Layered Beams", Int. J. of Mechanical Sciences, Vox. 21, pp. 489-504.
Pagano, N.J. and Pipes, R,B.‘ (.1971), "The Influence of Stacking Sequence
on Laminate Strength", J. of Con5)osite Materials, Vol. 5, January,
pp. 50-57.
Parkes, E.W. (1955), "The Permanent Deformation of a Cantilever Struck
Transversely at its Tip", Proc, of the Royal Society of London,
Vol. 228, March, pp. 462-476.
Parkes, E.W. (1958), "The Permanent Deformation of an Encastre Beam
Struck Transversely at any point in its Span", Proc. of the Institution
of Civil Engineers, Vol. 10, pp. 277-304.
Pipes, R.B. and Pagano, N.J. (1970), "Interlaminar Stresses in Composite
Laminates under Uniform Axial Extension", J. of Composite Materials,
Vol. 4, October, pp. 538-548.
103-15
Pipes, R.B. and Daniel, I.M. (1971), "Moire Analysis of the Interlaminar
Shear Edge Effect in Laminated Composites", J. of Composite Materials,
Vol. S, April, pp. 25S-2S9,
Radin, J. and Goldsmith, W. (1988), "Normal Projectile Penetration and
Perforation of Layered Targets", Int. J. of Impact Engineering,
Vol, 7, No. 2, pp. 229-2S9.
Ravid, M. and Bodner, S.R. (1983), "Dynamic Perforation of Viscoplastic
Plates by Rigid Projectiles", Int. J. of Engineering Sciences,
Vol. 21, No. 6, 1983, pp. S77-S91.
Ravid, M. , Bodner, S.R. and Holcman, I. (1987), "A Two-Dimensional
Engineering Model for Perforation of Layered Targets", Proc. of
the Tenth Int. Symposium on Ballistics, San Diego, Calif., October 27-29,
Volume 2, pp. 1-9.
Ravid, M. , Bodner, S.R. and Holcman, I. (1989), "Application of Two-
Dimensional Analytical Models of Ballistic Penetration to Ceramic
Armor", Proc. of the Eleventh Int. Symposium on Ballistics,
Brussels, Belgium, May, 3pp.
Recht, R.F. and Ipson, T.W. (1963), "Ballistic Perforation Dynamics"
ASME J. of Applied Mechanics, September, pp. 384-390.
Shivakumar, K.N. , Elber, W. and Illg, W. (1985a), "Prediction of Low-
Velocity Impact Damage in Thin Circular Laminates", AIAA Journal,
Vol. 23, No. 3, March, pp. 442-449.
Shivakumar, K.N., Elber, W. and Illg, W. (1985b), "Prediction of Impact
Force and Duration due to Low-Velocity Impact on Circular Composite
Laminates", ASME J. of Applied Mechanics, Vol. 52, September,
pp. 674-680.
Sodha, M.S. and Jain, V.K.. (1958), "On Physics of Armor Penetration",
J. of Applied Physics, Vol. 29, No. 12, December, pp. 1769-1770.
Taylor, G.I. (1948), "The Formation and Enlargement of a Circular Hole
in a Thin Plastic Sheet", The Quarterly Journal of Mechanics and
Applied Mathematics, Vol. 1, pp. 103-124.
Thomson, W.T. (1955), "An Approximate Theory of. Armor Penetration",
J. of .Applied Physics, Vol. 26, No. 1, January, pp. 80-82 and 919-920.
Woodward, R.L. (1978), "The Penetration of Metal Targets by Conical Projectiles",
Int. J. of Mechanical Sciences, Vol. 20, pp. 349-359
Woodward, R.L. (1987), "A Structural Model for Thin Plate Perforation by Normal
Inroact of Blunt Projectiles", Int. J. of Impact Engineering, Vol. 6, .No. 2,
pp’. 129-140.
103-16
Appendix A Extension of Thomson’s paper to spherical projectile
For a spherical projectile,, we have.
The aim of this appendix is to show that for a spherical projectile,
the petal ling equations provided by Thomson yields infinite work. This
shows that the spherical projectile is actually a "blunt -nose" projectile
and theoretically, plugging is the only possible perforation mechanisms
and petalling is theoretically not possible, using Thomson’s theory.
103-17
From Pythagoran theorem.
(R- Vt)^ + y2 = r2
solving for y, we have,
y2 = r2 _
that is,
y » (2RVt - V^t2)l/2
Taking the first derivative, one obtains,
y.^ » (l/2)C2RVt - (2RV - 2v2t)
that is.
Taking the second derivative, one obtians,
y,^^ = (-1/2) (2RV -V^t^)''’'^“(-2v2t) - (2Ry- vV)"^/^
103-18
x^+ y2 _ r2
y = Cr2-
= (1/2) (r2-x2)'^''^ C-2x)
dx
so that.
y. = (-X) (r2-x2)-i/2
^'xx ’ ■ + Cx/2) (r2-xV^^^ C-
2x)
that is,
y, » -(r2-x2)-1/2 - (x2) (r2-x2)-V2
Xx
The integrals are,
‘i= I
y y.tt
note that, dy _ dy dx _ y ^
dt ■ dx dt " ^
y,^^ = (y^t^’t = '^(y^t^’x ^
XX
dy = ^ dx
dx
x=0
x=0
Ij = V
>''xx ’'’X <*=' , 21^ = 2V^ J y (y.jj/ dx
x=R x=R
103-19
so that
x=0
Il+2i2 * V"
J'-xx >'*X " 2y ty.,)" dx
x=R
x»0
» J C-x) [-1 - (x^) Cr2-x2)-1] + 2 (-x^) (r2.x2)
x=R
x*0
» I X - (x^) (R^ dx
x»R
2 2
C-V R^/2)
XaO
x»R
dCx^)
Q =
2
-X ,
so that, x^ = R^ - Q
d(x
so that,
Ij + 2I2 = (-v2r2/2) - CV^/2)
Q=0
^ - 9,) (-1) dQ =
Q /
Thus,
II + 2I2 = C-V^R^/2) +
-1 dx
) = -dQ
infinite
103-20
1990 USAF-UES SUMMER FACULTY RESEARCH PROGRAM
Sponsored by the
AIR FORCE OFFICE OF SCIENTIFIC RESEARCH
Conducted by
Universal Energy Systems, Inc.
riNALREPQRT
ACCELERATE FATIGUE TEST PROCEDURE FOR THE
SiajgimL, PQLYgARBQNATE WMPQNEMI-Qf!. IHE
F-16 CANOPY COMPOSITE MATERIAL
Prepared by:
Academic Rank:
Department and
University:
Research Location:
USAF Researcher:
Date:
Contract No:
Yulian E. Kin, Ph.D.
Associate Professor
Engineering Department
Purdue University Calument
Flight Dynamics Laboratory
Robert E. McCarty, Lorene V. Garrett
1 August 1990
F49620-88-C-0053
by
Yulian B. Kin
ABSTRACT
The long-term fatigue test procedure requires the
breaking of 20 to 30 identically prepared specimens and one
month to complete. Thus, manufacturers often do not perform
a conventional fatigue test in spite of its obvious utility.
Therefore, there is a definite need for an accelerated
fatigue test which can be completed in approximately one day.
The accelerated test procedure proposed can be developed on
the basis of the data gained by the principal investigator
during the conventional fatigue test run with the help of the
UES mini-grant S-210-9M6-038 in 1989. The mini-grant was
awarded to continue the research started by Yulian Kin during
his summer appointment at Wright-Patterson Air Force Base in
1988.
104-2
ACKNOWLEI
I wish to thank the Air Force Systems Command and the Air
Force Office of Scientific Research for the sponsorship of
this research. Universal Energy Systems must be mentioned
for their concern and help to me in all administrative,
directional and social aspects during the summer appointment
and the pre-summer visit.
I am grateful to Group Leader Robert McCarty and Aerospace
Engineer Lorene Garrett. They provided support, attention,
and a comfortable working atmosphere. I appreciate very much
Lorene ‘s help with my computer needs. Many thanks to Mrs
Evelyn Schutte for the help in preparation of this report.
Dr Arnold Mayer's interest in every phase of my summer and
possible future projects served as a source of encouragement.
104-3
I. INTRODUCTION;
There are assumjptiohs based on preliminary tests that fatigue
mechanical properties of the polycarbonate sheets are
significantly vary from sheet to sheet ahd probably along the
same sheet.
Therefore, it is very important to have a vehicle which
permits to control (quickly and constantly) the quality of
the polycarbonate sheets and detect possible deviations of
the manufacturing process, in other words to control the
stability of the manufacturing process.
The accelerated fatigue test can be utilized for such kind of
control and the procedure of the test can be developed if
some fatigue parameters for the tested materials dr real
parts are known from the conventional test.
The accelerated fatigue test can be also useful to make a
preliminary estimate of new designs. For example, Flight
Dynamics Laboratory is initiating the investigation of the
fatigue characteristics of polycarbonate sheets with
different thicknesses.
For the obvious economical reasons a conventional fatigue
test should be run to confirm the expected possible output
only if the accelerated test shows the strong dependency of
fatigue strength on the polycarbonate sheet thickness.
1044
II. OBJECTIVES OF THE RESEARCH EFFORT:
The immediate objectives of the research were to:
1. Develop a procedure including a regime program of the
accelerated fatigue test for the coupons cut from a
structural polycarbonate sheet of the F-16 canopy composite
material. The necessary fatigue parameters (fatigue
strength, knee points and fatigue curve slopes for 5%, 90%
and 95% probability fatigue curves) were computed from the
statistics determined by Yulian Kin during his SFRP in 1988
followed by a mini-grant program in 1989.
2. Manufacture specimens in accordance with the procedure
given in (1] .
3. Start a conventional fatigue test using the procedure
given in [1] . It was my intention to complete 40% of this
test during the summer appointment and to continue at Purdue
University with funding from the mini-grant program.
4. Run (if time permits) a preliminary accelerated fatigue
test using the procedure from item 1 and statistics from item
3 and from [1].
III. LOCATY'S METHOD FOR ACCELERATED DETERMINATION OF
FATIGUE STRENGTH
1. Description of the Method.
The accelerated technique for the specified type of
specimens can be reliably developed if the statistics of a
conventional fatigue test of similar specimens are known.
Therefore, the results of the preliminary run conventional
fatigue test were used to assign certain parameters and
estimate the error of the accelerated test.
Locaty's accelerated technique was used to start an
investigation. It was assumed that time to run the proposed
accelerated test would be vithin 10 hours.
The idea and description of the method had been given in
appendix to [1] and repeats here in concise form for the
convenience.
The method is based on the concept of cumulative fatigue
damage or Palmgren - Minor rule considering ^
where n is the number of cycles which specimen worked in
the specified test regime, and N i is the number of cycles
which specimen could potentially work according to the
fatigue curve received from the results or the conventional
fatigue test of the same type of specimens.
J04-6
The loading program and the treatment of results are
presented in Figures 1 and 2. Figure 1 shows three fatigue
curves (for example, 5%, 50%, and 95% probability of faHure)
received from a conventional fatigue test.
During an accelerated test the specimen works, for example,
for 50,000 cycles at the first load level, then the load is
increased and the specimen is again tested for 50,000 cycles.
The procedure is repeated until the instant when the specimen
fails. Then using Figure 1, the magnitudes of
are determined. With these three parameters (or, if
necessary, a greater number of points) and knowing the
corresponding stresses, we can find the coordinates of the
points which result in the curve shown in Figure 2. Now, if
according to an accepted hypothesis fatigue strength
corresponds to a definite value (for example, I-
we can easily determine the fatigue strength magnitude
(Figure 2) .
104-7
ACCELERATED TEST
Fig 1. Loading Program Fig 2. Graphical Determination
of Fatigue Strength
2. Nomenclature and symbols
Parameter
Symbol
Unit
Refereice
Initial stress
C^o
Pa
Figure .1
Stress Increment
A(f
Pa
Figure 1
Initial number of cycles
Uq
Cycle
Figure 1
Number of cycles at ith
level of stress
Cycle
Figure 1
JL
Rate of stress increment
Pa /Cycle
Rate of stress increment
ik
Pa/ Cycle
Stress at failure
Pa
Fatigue strength
determinated by acceler.
method
Pa
Number of cycles at ith
level of stress determined
during conventional fatigue
test Cycle
Sum of relative lives
Fatigue strength
corresponding to curve A, p.
B, C respectively; S ;
Figure 1
Figure 2
Expected fatigue strength
Knee points (if any) on
curves h, B, respectively
Slopes of curves A, B, C
respectively
(fg,p. j Pa
Figure 1
^9 - N£xp
Kft
Kc
3 . Loading Regimes
Intervals of the initial load level are assigned from the
inequality: < (fa 0.2
Magnitude or j /^Iffxp^and Kg, Kq are assigned
after the completion of the conventional test. Optimal value
of the stress increment rate depends on the selected ratio
and can be (as in the first try) selected from 0.7 kpsi
X 10"^ to 3.5 kpsi X 10“® per cycle.
Optimal intervals for the stress increment are found from the
104-9
expression 0, C S ^ (f < ^.0 C^x/a
Value of n is determined from n * A C*
In case when n i appears to be greater than 3X10^ , it is
necessary to decrease A (f •
Most of the recommendations are based on test experience with
metal and corrections are quite possible.
The specimens, parameters and regimes selected for the
preliminary accelerated test are given below.
SPECIMENS
Specimen dimensions and loading diagram are given in Figures
Figure 3. Specimen Used in Study
Figure 4. Loading Diagram
104-10
REGIMES OF THE CONVENTIONAL TEST
Loads and stresses assigned for the conventional test are
given in Table 1.
Table 1.
Conventional Test Regimes
Max.
Load
on the
Machine
Gauge 1
Lc
Min.
Load
on the
Machine
Gauge
Lb
Amplitude
Load
Applied
to the
Specimen,
Lb
♦Nominal
Amplitude
Stress
Applied
to the
Specimen,
psi
Frequency,
HZ
Rmin
Rmax
1500
300
600
7200
3
0.2
1080
200
440
5280
4
0.2
720
144
288
3456
6
0.2
360
72
144
1728
12
0.2
180
50
65
780
15
0.28
Distance "a” (see Figure 4) is a constant for all
regimes .
a
♦Actual magnitude of amplitude stress calculated taking into
account actual specimen dimensions.
104-11
60p0;{l/2) (3/4)2
= 0..75
3(750)
4320(1/2) (3/4)
= 0.75
3(540)
2880 (1/2) (3/4)
- = 0.75
3>(360)
1440(1/2.) (3.4)
= 0.75
3(180)
50% probskbility fatigue curve ''B”(Fig. 5)15 built using
regression analysis of the data gained in< (I") fnr the,
specimens, with stress concentrators.
104-12
AMPLITUDE STRESS
Nimber of Cycles to Failure
Figure 5. 5%, 50%, 95% probability of failure fatigue
curves .
Slope of curve "B” K3 » 5
^ _ _
^ log N - log N
Expected fatigue strength for the life Ng® = 10® cycles
is (j = 1440 psi (selected from [1].
PARAMETERS FOR CURVES "A" and ''C
nJ = 0.8 No
ft A
1.2 No or No
6 c
No = No
Kc
Kb - 3
or K, - Kb
Kb + 3
0.9 (T^
Kc
INITIAL LOAD. LOAD INCREMENT. NUMBER OF CYCLES AT ONE LEVEL.
Initial Load:
Select initial amplitude stress = 1500 psi hence amplitude
load on the machine gauge is
^bh^ 1500(1/2) (3/4)2
- - - 18 7.5
3a 3(3/4)
from equations
Pmax - Pmin
and
£min = 0.2
Pmax
104-14
Pmax = 500 lb
Pmin = 100 lb
Load increment:
0.05 ( pj) < P < 0.20 (P^)
Pg« 450 lb (Table 1)
0.05 (450) < A P < 0.20 (450)
22.5 < AP <90
We assume that A P 100 lb
Number of cycles ni at one level
A
ni =
A C^max corresponding to AImax* 100 lb
(6.75) (3) (100)
- = 800 psi
(1/2) (3/4)2
If we take = 1.6 X 10"^ (see page 10)
0.8
ni - - 5” = 50000 cycles.
1.6 X lo”
IV TEST RESULTS
The MTS testing machine was calibrated just before the
test. The installed load magnitudes were red out from the
oscilloscope and additionally (for verification purpose)
controlled with digital multi-meter. 50% of the test was
completed and results are. given in Table 2.
Test Results
Table 2. Comparison of 1989 and 1990 Tests
Amplitude
Number of
Number of
Number of*
Stress ,
cycles to
cycles to
cycles to
psi
complete
crack
complete
failure,
initiation,
failure.
1990 test
1990 test
1989 test
7200
1200
1000
5280
4800
110000
3456
12800
8500
16600
11000
180000
23500
18000
17000
10500
1728
134300
85000
310000
156000
90000
960
1000000
450000
(No failure)
*The numbers were taken from Figure 4 [1].
From the results gained it can be assumed (take note that
the test has not been completed yet) that the 1989 materiel
had greater fatigue strength than the 1990 polycarbonate
sheet. But the scatter is significantly less in the 1990
test. The possible difference in strength of two
polycarbonate sheets of the same type shows one more time
that the tight and quick control of fatigue characteristics
is necessary to verify the product quality. In the same
time, relatively consistant repetition of the test results
can be used as a confirmation of the possibility to develop a
reliable accelerated test procedure.
V RECOMMENDATIONS
1. Continue the investigation to develop an
accelerated fatigue test procedure for the structural
polycarbonate of the F-16 canopy composite material. With
the help of mini-grant and some additional funds the work can
be completed in 1991.
2. Apply the accelerated testing technique to the
coupons with and without stress concentrators.
3. Investigate yhe dependency of fatigue
characteristics on thickness of the polycarbonate sheet. If
the dependence exists, interesting proposals to improve
design can be done.
104-17
4. Prepare a procedure for the fatigue test of the
canopy composite material. Preliminary hypothesis of failure
can be developed on the basis of finite element analysis.
5. Prepare a procedure for the environmental fatigue
investigations of the canopy material.
6. Fatigue investigations are very time consuming and
expensive. In the meantime, it is clear that Aircrew
Enclosures Group of Vehicle Subsystems Division needs to
provide different fatigue investigations in the near future.
Taking this into consideration. It is proposed to establish
a mutual research laboratory with the installation of a new
modern fatigue machine at Purdue in Hammond. The testing
machine which vehicle Aircrew Enclosures Group is planning to
buy can be partially paid from the Purdue funds. Purdue
University Calumet Engineering Department has space for
installation and experienced personnel to provide
investigation, maintenance and calibration of the machine.
Engineering senior students will handle the tests while
developing their senior design projects. (Graduate students
can be also attracted.) For the first five years the machine
will be used for the Aircrew Enclosures Group projects only.
After five years the machine can be returned to Wright-
Patterson or continue to be installed at Purdue with the
availability for other projects also. The economical
benefits of the proposal are obvious (;the expenses are
104-18
divided between two institutions working on the same
projects, the time of tests is decreased because of the
number of students available for the work, the student salary
is much less than for the regular personnel) . The
educational benefits are also very significant because the
students will get the opportunity to work on very serious,
interesting and practical projects. The principal
investigator has already in general discussed this proposal
with the department head and met his understanding.
104-19
1. Yulian Kin., "Fatigue characteristics of F-16 composite
transparency material determined by long-tera and accelerated
methods," Final Report, Contract No. FY9620-88-C-0053/SB5881-
0378, Universal Energy Systems, Inc., Dayton, Ohio, 1990.
104-20
1990 USAF-UES SUMMER FACULTY RESEARCH PROGRAM
GRADUATE STUDENT RESEARCH PROGRAM
Sponsored by the
AIR FORCE OFnCE OF SCIENTIHC RESEARCH
Conducted by the
Universal Ener^ Systems, Inc.
FINAL REPORT
Study of Fracture Behavior of Cord-Rubber Composites for
Lab Prediction of Structural Durability of Aircraft Tires
Prepared by: Byung-Lip ("Les") Lee
Academic Rank: Associate Professor
Department and
University:
Research Location:
USAF Researchers:
Engineering Science and Mechanics .
Pennsylvania State University, Univ. Park, PA
Wright-Patterson Air Force Base
Wright Research and Development Center
Flight Dynamics Laboratory
Vehicle Subsystems Division
Aircraft Launch and Recovery Branch
Gary J. Migut, John P. Medzorian
Date:
30 September, 1990
F49620-88-C-0053
ContractNo:
Shidv of Fracture Behavior of Cord-Rubber Composites for
I. ah Prediction of Structural Durability of Aircraft Tires
by
Byung-Lip ("Les") Lee
ABSTRACT
An aircraft tire durability study is underway to investigate the deformation and fracnire
behavior of cord-rubber composites. This study will identify the important parameters
responsible for the structural failure of aircraft tires by the use of analytical and laboratory
prediction methods, These methods will also identify the interaction between material
property degradation and damage accumulation in cord-rubber composites. Preliminary
results using coupon specimens of tire carcass have revealed that prolonged static and
cyclic loading sequences produce extensive interply shear deformation at the free edges
resulting in cord-matrix debonding followed by delamination type failure. These loading
sequences represent the circumferential tension in the footprint region of aircraft tires. It
was also determined experimentally that ^ fatigue endurance limit can be established for
cord-rubber composites. Analytical methods using finite element models of coupon
specimens have demonstrated reasonable accuracy in predicting load-displacement response
and interply shear strain variations. Future plans will include the correlation between the
fatigue resistance data of composite specimens and dynamometer test results of actual tires.
105-2
I wish to thank the Air Force Systems Command and the Air Force Office of Scientific
Research for sponsorship of this research. Universal Energy Systems helped me in all
administrative aspects of this program.
I sincerely appreciate continuing support and encouragement for this research work from
IVtessrs. John P. Medzorian, Gary J. Migut, Paul M. Wagner, Paul C. Ulrich and Aivars
V. Petersons, Mrs. Monika H. Champion and Dr. Arnold H. Mayer at the Vehicle
Subsystems Division of WRDC Flight Dynamics Laboratory. Along with ixiy graduate
student Mr. Patrick M. Fourspring (Penn State Univ.), Messrs. John P. Medzorian, Gary
J. Migut and Paul M. Wagner participated this research program as co-investigators.
I am also very grateful to Dr. George P^ Scndeckyj (WRDC Flight Dynamics Laboratory,
Structures Div,), Dr. Ran Y. Kim (Univ. of Dayton Research Inst.) and Dr. James M.
Whimey (WRDC Materials Laboratory, Non-Metallic Materials Div.) for their support of
the research program. Mr. Harold Stalnakcr (WRDC Flight Dynamics Laboratory,
Structures Div.) and Messrs. Ron Esterline md John Camping (Uniy. of Dayton Research
Inst.) played an essential role in this program by performing various mechanical testing.
Finally my sincere thanks should go to Drs. Alfredo G. Causa and Yao M. Huang
(Goodyear Tire & Rubber Co.) and Dr. Chong K. Rhee (Uniroyal Goodrich Co.) for
providing composite specimens.
1. INTRODUCTION
Compared with other types of pneumatic tires, aircraft tires are subjected to unusual
combinations of speed and load (/). For example, the baseline tire used in this smdy is a
49X17/26PR bias constmetion and is rated at a speed of 358 km/hr (224 mph) and a load
of 173 500 N (39 000 Ibf). Extreme combinations of speed and load result in high cyclic
frequencies, large deflections and significant heat generation due to hysteretic loss in
aircraft tires. As confirmed by field experience, these conditions cause damage in critical
sub-regions of tires such as shoulder ^ bead, lower sidewall or tread which eventually
develops into catastrophic failures of whole tires. The failure of the carcass ply in the
shoulder area, often called a ply separation, occurs in the form of delarhinatidn which
involves crack propagation mainly in the rubber matrix and cord-matrix interface (Figure
1). So-called bead area cracking and lower sidewall break also involve crack propagation
in the rubber matrix of carcass ply, but there are strong indications of fiber fracture as well.
The processes of damage accumulation and structural failure of tire carcass in the shoulder,
bead and lower sidewall areas are attributed to a combination of mechanical overloading
and heat generation along with the resultant deterioration of constituent materials.
On the other hand, tread failure of aircraft tires is attributed to the presence of strong
centrifugal force resulting from unusually high speed and high load requirements. Tread
failure which starts in the form of circumferential cracks at the base of tread groove is
caused by repeated groove flexing motion when so-called standing wave is present.
Although the term derives from their stationary appearance to the observer, standing waves
occur when the natural circular frequency equals a critical frequency (i.e. critical speed)
with which a section of tire passes the ground contact position (Figure 2). The repeated
deformation due to standing waves produces a significant heat build-up and thereby
deteriorates material properties. When circumferential groove cracks caused, by standing
waves expose cord reinforcement of underlying c^cass ply, they develop into a complete
tread-carcass separation ratlier quickly leading to catastrophic failure of tires. Past research
works (1-3) showed that rotational speeds of tires approaching critical speed should be
avoided to prevent premature tread failure.
As reviewed so far, basic understanding of tire failure mechanisms has been established at
least in qualitative sense. In the case of tread failure, its major cause, the occurrence of
standing waves can be avoided by a pto^ design of tires (1-3). However, as fa- as the
carcass ply is concerned, the tasks of identifyng critical operating conditions (speed, load,
underinflation and tire deflection) responsible for the failure and predicting the useful life
expectancy of an aircraft tire are still difficult at best. Accelerated testing based on
dyn^ometers provides a valuable means of evaluating the structural durability and life
expectancy of aircraft tires. But the results of these dynamometer tests reflect merely the
sensitivity of each specific tire design and construction to a specffic combination of test
conditions, unless underlying mechanisms of property degradation and structural failure are
identified.
Past approaches (4-6) to characterize the property degradation of the aircraft tires have
utilized coupon specimens cut from the carcass region. Clark et al (4) conducted a smdy
which assessed the degradation of adhesion strength between the carcass plies of
conditioned tires through the peel test of coupon specimens. His study also examined the
timenemperature dependence of peel strength. Another smdy (5) is currently attempting to
develop experimental procedures which affect the degradation of adhesion strength of the
tire coupon specimens by subjecting them to preload, flexural fatigue and temperature
conditioning. The durability analysis of Ref provides useful infonnation on failures
caused by the deterioration of material properties such as ply-to-ply adhesion strength.
Some estimates assume that dre failure occurs' at a SO percent reduction’ of carcass ply
adhesion strength. However, these works did not show how the deterioration of material
properties leads to damage accumulation and eventual structural failure of tire carcass. The
research program de^ribed in this paper has begun and plans to achieve the laboratory
prediction of the durability of aircraft tires based on the smdy of deformation, damage
accumulation and ftaemre behavior of ewd-rubber composites.
n. OBJEimVES OF THE' RESEARCH EFFORT
The study will investigate the mechanisms of carcass ply failure of aircraft tires with an
emphasis on failure modes in the shoulder area. Appropriate test methodologies will be
chosen to simulate ply delawination using cord-rubber composite specimens. By assessing
inteipiy shear deformation behavior of composites under static and fatigue loading, the
smey will identify the important parameters responsible for the damage accumulation and
stnctural failure. This assessment will also define the interaction of damage accumulation
and material property degradation. Analytical modeling efforts will be performed for the
105-5
predction of interply she^ strain in tire carcass composite specimens. The study will
correlate the laboratory data of fatigue resistance of cord-nibber composites with the
durabili^ test results of actual aircraft tires.
m. EXPERIMENTAL STUDY OF COMPOSITES
In developing test metliodologies for the simulation of carcass ply delamination type
failures, the study used two kinds of lab coupon specimens press-molded from the
calendered ply stocks. The first type of specimen consists of a cord-rubber composite
simulating typical bias aircraft tire carcass using a +/- 38 deg reinforcement angle, 1260/2
nylon cord and a proprietary rubber compound based matrix. The second type, which is a
model composite, uses a +/- 19 deg cord angle, steel wire cables of circular cross-section
and a natural rubber-based compound matrix. The model composite coupon specimen was
included in the study to allow a better determination of the failure modes, because the large
cord diameter and the cord angle chosen maximize the interply shear strain, a major
contributing factor in composite failure (7),
To avoid tension-bending coupling, the specimens were constructed with a symmetric ply
lay-up. The end tabs were added to the specimens to prevent failure near the gripping
region. All other specifications of composite coupon specimens are listed in Tables 1 and
2. To represent the circumferential tension of an aircraft tire in the footprint region, coupon
specimens were subjected to both static and cyclic uniaxial tensile loading. Cyclic testing
was performed under a broad range of load amplitude and at three different levels of
frequency (speed) with a close monitoring of heat generation. In both static and cyclic
testing, local strain was estimated by measuring the displacement of line markings drawn
on the specimen edge as shown in Figures 3 and 4.
Under both static and cyclic tension, the composite specimens were found to exhibit a large
interply shear s/ram (Figures 3, 5 and 6). As analyzed by numerous studies in the past (7-
9), intei-ply shear strain develops in angle-ply laminates when the constituent plies exhibit
in-plane shear deformation of opposite direction but the action is prevented by mutual
constraint due to interply bonding. Compared with the case of fiber-reinforced plastic
composites, cord-rubber composites exhibit unusually high level of interply shear strain
which results from the load-induced change of reinforcement angle allowed by extreme
compliance of mbber matrix. The results of the relative displacement for two adjacent plies
105-6
show that, at an axial strain of 10 percent, an interply shear strain of 70 percent develops in
the nylon cord-reinfoiced composites and 120 percent in the model composites as shown in
Figures 5 and 6.
Above a critical value of interply shear (static) strain corresponding to approximately one-
thirtl of ultimate strength of the composites, localized^failure is induced in the form of cord-
matrix debonding. Debonding is started around the cut ends of reinforcing cords at the
edge of the finite width coupons as shown in Figures 3 and 4. This phenomenon is
justified since the maximum interply shear strain occurs at the edge of the specimen. The
role of cut ends of cords as built-in defects in the process of damage initiation confirms
detailed observations made by Breidenbach and Lake (10,11). The model composites,
shown in Figure 4, reveal the progressive development of cord-matrix debonding into
matrix cracking as the strain increases. At higher strain, the axial stress-strain curves for
both nylon cord-reinforced and model composites exhibit strain hardening type response
(Figures 7 and 8).
A new finding was that the above-described critical load for the onset of cord-matrix
debonding constitutes a threshold level for semi-infinite fatigue life, i.e. fatigue endurance
limit, of the composites. As shown in Figure 9, the S-N curve (maximum cyclic stress vs
the number of cycles to failure) for the model composites has a distinct endurance limit.
Preliminary results also indicate the same behavior for nylon cord-reinforced composites.
The physical description of an endurance limit is that, with cyclic stresses lower than the
limit, cord-matrix debonding and matrix cracking are never initiated nor developed. Under
cyclic stresses exceeding the endurance limit, debonding and matrix cracking are
progressively widened and developed into the delamination leading to gross fracture of the
composites (Figure 4).
The observed process of damage accumulation was accompanied by a continuous increase
of temperature (and cyclic strain as shown in Figure 10. The stabilization of temperature or
strain (zero rate of increase) was determined to be a good indication of infinite fatigue life
(Figure 11). Either by degrading cord-matrix adhesion strength or by lowering the
modulus of rubber matrix, the increase of temperature may contribute to the increase of
cyclic strain. The increase of cyclic strain at a given stress amplitude will in turn produce
more hysteretic heating which can further reduce fatigue lifetime of tire carcass composites.
However, the possibility of this type of interaction involving the changes of cord-matrix
adhesion strength or matrix modulus seems to be rather remote, since the measured rates of
105-7
temperature and strain increase are almost linear (Figure 10). Additional fati^e
experiments of cord-rubber composites at elevated temperatures or after heat aging are
planned to resolve the question.
IV. COMPUTER MODELING
Experimental study has shown that cord-rubber composite coupon specimens representing
the aircraft tire carcass exhibit extensive interply shear deformation between constituent
plies under circumferential tension, which evenmally leads to delamination type failures. In
order to investigate the variation of interply shear strain within composite specimens, a
finite element model (FEM) using the ADINA computer code was created by co¬
investigator J. P. Medzorian (WRDC Flight Dynamics Laboratory, Vehicle Subsystems
Div.). Linear elastic-onhotropic material elements were used to model the constitutive
behavior of the cord-reinforced plies, while linear elastic-isotropic material elements were
used for modeling of rubber layers. Since shear strains vary linearly throughout the solid
finite elements, the model is restricted to only predicting linear variations of interply shear
strain through the coupon thickness.
Under an uniaxial tensile load of 3 491 N, the axial displacement of the model composite
coupon was analytically predicted to be 4.06 mm which was within 14 perecent error from
the experimental value of 3.55 mm. The analytically predicted value of interply shear strain
of 80 percent compares well with experimentally measured values of 65 to 85 percent. In
the case of nylon cord-reinforced composite that simulates a bias aircraft tire carcass, the
axial displacement under a uniaxial tensile load of 792 N was predicted to be 21.26 mm,
which is very dose to an experimental value of 21.59 mm. The agreement between the
analytical predictions and experimental results was less satisfactory for the variation of
interply shear strain. Under the load of 827 N, the predicted value of maximum interply
shear strain was 13 perecent which is much less than experimentally observed values of 25
to 35 perecent. One cause of error could be mesh refinement. Since solid finite elements
are limited to predict only linear variation of interply shear strain, many more elements are
required in each layer to accurately predict non-linear shear strain variations.
V. CORRELATION WITH TIRE TEST RESULTS
105-8
To correlate the laboratory data of interply shear fracture resistance of cord-rubber
composites with durability test results of «;tual aircraft tires, an extensive tire test plan has
been developed jointly with co-investigators P. M. Wagner, G. J. Migut and J. P.
Medzorian (WRDC Flight Dynamics Laboratory, Vehicle Subsystems Div.); The test
program will use 49X17/26PR KC-135 bias aircraft tires. The durability tests will be
accomplished on a 120 inch diameter dynamometer located in the Landing Gear
Development Facility at Wright-Pattcrson Air Force Base. The tests will be performed with
a systematic variation of footprint load, irfladon pressuremd speed. For each variation of
conditions, carcass ply delamination as a dominant failure mode of tires will be confirmed
by fractography. The study will examine the effects of footprint load, inflation pressure
and speed on the mileage to failure (lifetime), deflection and temperature history. An
establishment of a threshold load for damage initiation is planned to determine an endurance
limit for the tires.
These results of tire durability will be examined in comparison with laboratory fatigue
resistance data of carcass composite specimens. The composite data will include the effects
of stress amplitude, ratio of maximum to minimum stress and cyclic frequency on the
number of cycles to failure, strain increase rate, heat-up rate and fatigue endurance limit.
By comparing how deforaiation and fracture properties lure dependent on the same
controlling parameters, the dynamometer test results of acmal tires will be correlated with
the fatigue resistance data of composite specimens. In addition to the parameter correlation
study, coupon specimens will be extracted from the tires with various load histories and
then subjected to static and fatigue tests to characterize the degradation of material
properties.
VI, CONCLUDING REMARKS
The described research effon aims at the laboratory prediction of aircraft tire durability
based on the study of deformation, damage accumulation and fracture behavior of cord-
rubber composites. As a major step toward the goal, the study plans to identify critical
operating parameters in terms of stress, strain and temperature history responsible for
structural failure of tire carcass. The smdy will also define the mechanisms of damage
accumulation and its interaction with material property degradation. Emphasis is placed on
carcass ply delamination'm the shoulder region of aircraft tires.
105-9
Experiment^ methods have been developed to Emulate delaminadon process in cord-
rubber composite specimens. Under static or cyclic loading which represents
circumferential tension in the footprint region of tires, angle-ply composite coupon
specimens were found to exhibit extensive interply shear deformation which eventually
leads to delaminadon type failures. The study determined that a fatigue endurance lirnit can
be established. The process of damage accumulation was always accompanied by a steady
increase of temperature and local strain.
The analytical model developed by co-investigator J. P. Medzorian was able to predict the
deformation behavior of cord-rubber composite specimens. The finite element model based
on the computer code ADINA accurately predicted axial load-displacement relationship and
the variation of interply shear strain through the thickness.
Finally a plan was developed jointly with co-investigators P. M. Wagner, G. J. Migut and
J. P, Medzorian to correlate the laboratory fatigue resistance data of cord-rubber
composites with dynamometer test results of actual aircraft tires. The dynamometer tests
will assess the effects of footprint load, inflation pressure and speed on the lifetime,
deflection and temperature history of the tires. The results will be compared with the data
showing the dependence of corresponding composite properties on the same controlling
parameters.
Vn. RECOMMENDATIONS
The following recommendations can be made for the future study on fatigue fracture
behavior of cord-rubber composites and laboratory prediction of aircraft tire durability:
(1) Using developed experimental and analytical methods, identify the critical parameters
of stress, strain and temperature history that are responsible for damage accumulation and
delamination failures of aircraft tire carcass composite specimens under fatigue loading
conditions. Some stress history parameters include the mean stress, stress amplitude and
frequency.
(2) Determine the importance of hysteretic heating, dynamic creep and degradation of
constituent material properties \n controlling the processes of damage accumulation and its
development into delamination failures of tire carcass composite specimens.
105-10
(3) In order to evaluate the degree of damage accumulation, it is necessary to perform a
detailed assessment of the fadguerinduced degradation of static strength and stiffness for
the tire c^ass composite specimens. A study is needed to determine if the fatigue lifetime
of cord-rubber composites is dependent on the rate of damage accumulation but
independent of damage history.
(4) Based on these experimental results, assess the applicabili^ and limitations of classic
approaches in the theoretical prediction of the fatigue lifetime of cord-rubber composites
under the various combinations of load amplitude and mean load (Goodman's) and
complex loading sequence (Miner's).
(5) Develop experimental methods to accurately measure the stress, strain and temperature
history in cord-rubber composites as well as aircraft tires during simulated operations.
Specifically, the measurement of interply shear strain and stress is needed.
(6) Develop test methodologies for the lab simulation of other types of failures processes
besides shoulder delamination of aircraft tires. Other common types of failure processes
include bead and lower sidewall area failure.
(7) Examine the effects of footprint load, inflation pressure and speed on the mileage to
failure, deflection and temperature history of aircraft tires and assess the failure modes of
tires based on fractography. Correlate these results with the fatigue resistance data of cord-
rubber composites.
105-11
References
(1) J. H. Champion and P. M. Wagner, "A Critical Speed Study for Aircraft Bias Ply
tires", AFWAL-TR-88-3006 (1988).
(2) J. H. Champion, S. K. Clark and M. K. Hilb, "A Study of Vibrational Modes in
RolUng Aira^t Tires", WRDC«TR-89-3092 (1989).
(J) J. Padovan, "On Standing Waves in Tires", Tire Science and Technology, Vol. 5, No.
2, p.83 (1977).
(4) S. K. Clark, "Loss of Adhesion in Cord-Rubber Composites", Report on Contract
No. 83-P-80607, U.S. Department of TrMsportation, Transportation Systems Center,
Cambridge, MA (1984).
(5) Nonh Carolina A&T State University, "Study of the Behavior of Aircraft Tire
Coupons under Various Losing Conditions", Report on Contract No. F33615-87-C-
341 1 U.S. Department of Air Force, Air Force Systems Command, Wright-Patterson
AFB, OH (1990).
(6) Personal communications with the researchers in tire industry.
(7) J. D. Walter, "Cord-Reinforced Rubber" in Mechanics of Pneumatic Tires edited by
S. K. Clark, U.S, Department of Transportation, Washington D.C. (1982)
(5) Interlaminar, Rcspgnsg of Compositt, Materials edited by N. J. Pagano, Composite
Materials Scries Vol, 5, Elsevier (1989).
(9) J. L. Ford, H, P. Patel and J. L. Turner, "Interlaminar Shear Effects in Cord-Rubber
Composites", Fiber Science and Technology, Vol. 17, p.255 (1982).
(10) R. F. Breidenbach and G. J. Lake, "Mechanics of Fracture in Two-Ply Laminates",
Rubber Chemistry and Technology, Vol. 52, p.96 (1979).
(11) R. F. Breidenbach and G. J, Lake, "Application of Fracture Mechanics to Rubber
Articles Including Tyres", Philosophical Trans. Royal Soc. London, Vol. A299, p.l89
(1981).
105-12
lAfiLLI
Soecificadons of Aircraft Tire Carcass Conposites
( Cord; 1260/2 Nylon cord )
( Matrix: Proprietary rubber compound)
Cord Angle
Cord Modulus
Matrix Modulus
Cross-Secdonal Area of Cord
Total Width of Specimen
Total Thickness of Specimen
Specimen Gage Length
-38, ^5S. +38. -33 deg
2.07 G?a (300X10^ psi)
5.51 .MPa (800 psi)
0.342 mm- (5.3.X10"^ inch4)
25.4 mm 1 1 inch i
6.35 mm (0.25 i.".ch)
101.6 mm (4 inch)
Specifications of Model Angle-Ply Composites
(Cord: Steel wire cable)
(Matrix: Natural rubber based compound)
Cord Angle
Cord .Modulus
Matrix Modulus
Cross-Sectional Area of Cord
Total Width of Specimen
Total Thickness of Specimen
Specimen Gage Length
-19. -19. +19. .;9deg
207 GPa(30XI(Ppsi)
5.51 MPa (800 psi)
2.06 ram- (3.2X10'-' inch-)
25.4 mm tl inch)
9.29 mm (0.37 :.".ch)
101.6 mm (4 inch)
105-13
Figure 2
Standing Waves in Aircraft Tires (Courtesy ofM. H. Chair^ion at Flight Dynamics Lab)
105-14
Figyre 3.
Cord-Matrix Debonding of Angle-Ply Composites Due to Interpiy Shear Deformation
(Model composites under tensile fatigue loading with 23 MPa maximum and 7 MPa
minimum stress: a- Before loading; b- Initial static loading; c- After 40 000 cycles; d-
After 460 0(X) cycles at 5 Hz frequency)
105-15
Figure 4
Other Failure Modes of Angle-Ply Composites Due to Interply Shear Deformation (Model
composites: a- Cord-matrix debonding into matrix cracking; b- Matrix cracking into partial
delamination)
•5
I
6
a
<
3
t-
T
T
T
6b
4b
2b
Bias Atrctaft Tire Composites
gross delamnation -
la
o aP
99
>S>
o Oq
10 ^50 30
AXIAL iSTRAlN (%)
5
•X3
x;
u
e
Q
<
3
10 eo;
CO
u
I
.4
<
O
2
->0
Figurei
Load vs Axial Strain for Aircraft Tire Carcass Composites (deflection rate = 5 nun/min)
2 r
Q
■<
o
2b
- 00—
o o oS
O (J,00
o oo
OCBO
0 oa
ocno
OOCP
•gross deUtmination
H2
CD
1®
Model Angle-Ply Composites
10
30
AXIAL STRAIN (%)
-llO
•5
•o
u
s
Q
<
O
CO
CO
ut
CO
1
o
2
-JO
Figure 8
Load Axial Strain for Model Angle-Ply Composites (deflection rate = 1.25 mm/min)
105-18
MAXIMUM CYCLIC STRESS (MPa)
MAXIMUM CYCLIC
Fiairc-LQ
Continuous Increase of Deflection and Temperature under Cyclic Loading for Model
Angle-Ply Composites (28 MPa max. stress, 7 MPa min. stress, 5 Hz freq.)
MAXIMUM cvaic STTtESS (ksi)
Figure 1 1
Rate of Temperature Increase vs Maximum Cyclic Stress for Model Angle-Ply
Composites
105-20
Prof. Vemoii Matzen
Technic^ Repoit submitted to Laboratory
106-2
1990 USAF-UES SUMMER FACULTY RESEARCH PROGRAM/
GRADUATE STUDENT RESEARCH PROGRAM
Sponsored by the
AIR FORCE OFFICE OF SCIENTIFIC RESEARCH
Conducted by the
Universal Energy Systems, Inc.
FIHAL.. BEFQRT
P.EIAMiyATIQM...a£..LAMlNATEg. CfiMPgSXTES
Prepared by:
Academic Rank:
Department and
University:
Research Location:
William E. Wolfe
Associate Professor
Civil Engineering Department
The Ohio State University
WRDC/FIBCA
Wright Patterson AFB
OH 45433
USAF Researcher:
Date:
Contract No:
Dr. Raghbir S. Sandhu
30 Sep 90
F49620-88-C-0053
DELAMINATION OF LAMINATED COMPOSITES
by
William E. Wolfe
ABSTRACT
In previous summer faculty research program appointments we
have looked at the initiation of damage in laminated
composites subjected to low velocity impact. A review of the
literature as well as an analysis of our own tests showed that
a significant mode of failure resulting from the impact event
is delamination. A prediction of the extent of delamination
requires an evaluation of interlaminar stresses and the
material properties governing delamination.
The research performed during this summer's appointment
followed two different lines. In the first effort, the
theoretical studies begun in a 1989 mini-grant to determine
the state of stress at each interface in a laminated composite
plate subjected to a dynamic load were continued. In the
second line of study, an analytical and experimental
investigation of the tendency for delamination as predicted by
the delamination moment coefficient originally defined by
Sandhu was performed.
107-2
Acknowledgements
I wish to thank the Air Force Systems command and the Air
Force Office of Scientific Research for sponsorship of the
Summer Faculty Research Program. Universal Energy Systems
Inc. administered the project, Their help throughout the
summer is appreciated.
My summer experiences with the Air Force research community
continue to be rewarding because everyone in the Flight
Dynamics Laboratory at WRDC is interested in making it so. I
am grateful to all who have helped me during this past summer.
I wish to give particular thanks to Mr George Holderby, group
leader of FIBCA for most of the past four summers, who made
the work environment so conducive to success and to Dr. Jim
Olsen, the Chief Scientist of the Flight Dynamics Lab, who has
somehow found the time to take a personal interest in this
research.
I am particularly grateful to Dr. Raghbir S. Sandhu who has
been my technical monitor in this and three previous summer
programs. Dr. Sandhu so enjoys sharing his huge supply of
research ideas that I am always disappointed to see the summer
come to an end.
107-3
I.
INTRODUCTION
The behavior of laminated composites that have been impacted
at low velocities is of great concern to designers of high
performance aircraft. These impacts which may result from a
maintenance accident or debris impact during takeoff or
landing can significantly degrade aircraft performance and
even cause structural failure without detectible surface
evidence of the event. A large number of researchers have
studied the delamination problem both analytically and
experimentally. In spite of considerable effort, current
predictive capabilities are limited. The research conducted
during this summer faculty appointment was a continuation of
an ongoing effort, begun with a 1987 Summer Faculty
appointment, to develop a predictive model for the response of
laminated graphite epoxy plates subjected to low velocity
impact (5,6). Theoretical studies begun by the author with a
1989 mini-grant (7), to determine the state of stress at each
interface in a laminated pomposite plate that results from the
application of a dynamic load were continued. In addition, an
analytical investigation of the tendency for delamination
using the delamination moment coefficient as originally
defined by Sandhu (2) as the delamination criterion was
conducted .
1074
II. OBJECTIVES OE THE RESEARCH EFFORT
The work described in this report is a continuation of
previous summer faculty research program appointments (5,6) in
which the initiation of damage in laminated composites
subjected to low velocity impact was studied. Those efforts
showed that, during the impact event, a significant cause of
failure is ply delamination.. , A prediction of the extent of
delamination requires an evaluation of both the interlaminar
stresses and the material properties governing delamination.
The main objectives of this summer's research program resulted
in two separate efforts. The first task was to cbhtinue the
theoretical studies begun in a 1989 mini-grant to determine
the state of stress at each interface in a laminated composite
plate that results from the application of a dynamic load.
The goal was to extend this work by deriving a self-adjoint
fonn of the governing equations. This would enable solution
schemes based on finite element methods to be developed. The
second task consisted of a numerical investigation of the
tendency for delamination as predicted by the delamination
moment coefficient originally defined by Sandhu (2) . The
objective of this second effort was to identify ply
orientations in which delamination was the sole mode of
failure so that test specimens suitable for use in determining
interlaminar strength properties could be designed.
107-5
III. RESULTS OF WORK EFFORT
3.1 Theoretical Studies:
There has been a considerable amount of research conducted
over the past three decades on the behavior of composite
materials. Much of the work has been in the area of laminated
plate theory which has, in recent years, focussed on ply-by-
ply analyses. Reviews of previous work have been presented by
several investigators including the author and his students
(Wolfe and Schoeppner (7), Schoeppner (4) ) , and the SERF
technical monitor of this project (Sandhu (2), Sandhu and
Sendeckyj (3) ) . Wolfe and Schoeppner discussed the advantages
of a stress based theory for describing the behavior of
laminated composites. In Pagano's stress based theory (l),
a linear variation of in-plane stresses is assumed. The
transverse stresses are obtained by integrating the
equilibrium equations. This theory has been shown to
accurately model the high stress gradients near the free edge
of a laminated plate while maintaining stress continuity at
the ply interfaces. However, the stress based formulation as
presented by Pagano does not allow for inertia effects. The
extension of this method to include dynamic behavior has been
the subject of the present study since such a method would be
a valuable tool in the determination of stresses in laminated
composites subjected to transient loading, in a recent report
to UES, Wolfe and Schoeppner (7) described the initial steps
in the development of an extension of Pagano's method to
IG7-6
include the effects of the inertia terms in the expression for
the ply-by-ply stress field. During the SFRP ah energy
formulation was used to derive the governing equations. A
self-adjoint system of the governing equations was obtained
consisting of twenty-three field equations. The derived
formulation is presented in more detail in Schoeppner (4) .
3.2 Preliminary Numerical Studies;.
Sandhu (2) in a review of a number of theoretical and
numerical programs, reported that most investigators have
identified interlaminar normal stresses as the primary cause
of ply delamination. He observed that a number of attempts
had been made to evaluate the magnitude of the interlaminar
normal stresses due to inplane loads. In several of these
studies expressions for delamination tendency were based on an
assumed distribution of interlaminar normal stresses. In his
work, Sandhu observed that these assumed distributions of
interlaminar normal stresses were not necessary for
determining the ply orientations most prone to delaminate.
Sandhu defined a delaminating moment coefficient (DMC) ,- and
showed that delamination could be expected when the DMC
reached a certain critical value. This delaminating moment
coefficient given in teinns of the average axial stress is
presented as equation 1.
107-7
DMC^=^^m{Am~h) (-4^) —
O
jfiq'. 1
where:
DM = delaminatihg moment at the mid-surface of the
laminate
m = number of 90° plies
n = number of 0° plies
t a ply thickness
= transverse stress in the '0 plies
0 o ^ average axial stress
Subsequent work by Sandhu and Sendeckyj (3) identified the
critical value of the DMC as being a function of material
type, plate thickness and ply orientation. They observed that
laminated graphite/ epoxy composites with DMCs greater than
10x10’® in® delaminated, whereas plates with ply orientations
resulting in DMCs less than 8x10’® in® did not delaminate. They
speculated that the DMC could be considered to be a material
parameter capable of defining the delamination tendency of
composites of different material systems.
An analysis of the effects of different ply orientations on
the value of Sandhu 's delamination moment coefficient was
performed for plates of several thicknesses with off-axis ply
orientations varying from 0 to 90 degrees. Figure 1 shows the
results of these calculations. It is apparent that, for the
graphite/ epoxy system studied, a critical angle for the off-
axis plies, between 40 and 50 degrees can be defined. It is
also clear from Figure 1 that the DMC and therefore
delamination tendencies increase with increasing laminate
thickness. Saridhu and Sendeckyj have shown that although
coupons made with the angle of the off-axis plies slightly
greater than 40 degrees delaminated, they also experienced
axial strains greater than 4000 uin/in prior to the onset of
delamination. Since the failure strain of the epoxy matrix is
approximately 4000 uin/in, any delamination in these specimens
would be associated with matrix cracking and not just a simple
opening mode failure. Therefore it was determined that for
the finite element studies to be conducted in this program,
only laminates with the off-axis plies oriented at 40 degrees
to the long axis of the test coupon would be exanined.
3.3 Finite Element Studies:
Using the results of the above preliminary numerical studies,
a finite element analysis of the specimens with the greatest
tendency to delaminate without experiencing natrix cracking
during in-plane loading was undertaken. In tliis section, we
present the results of these studies which were conducted to
identify the state of stress and thereby the tendency for
delamination in graphite/epoxy plates of six different
thicknesses. As was done in the preliminary studies, each
plate was modelled as having two 90 degree plies located at
the midplane of the specimen. The outer plies all consisted of
fibers oriented at ±40 degrees. The number of pairs of +40
107-9
degree plies placed syminetrically about the 90 degree plies
ranged frbm one, [±45/90]5, to six, [ (±45),^90]j. The results
of the computer simulations of the tensile test for the
thinnest coupon studied [±45/90]^, and the thickest,
[(±45)^90]j, are shown in Figures 2 and 3. Note the clear
tendency to peeling as well as stretching in each of the
figures. The stress contours plotted for each of the coupons
studied shows that the interlaminar normal stresses are
considerably larger and the gradients steeper for the thicker
specimens than for the thinner ones, thus verifying the
finding of Section 3.2 that the likelihood of delamination
increases with the thickness of the panel. The distribution of
interlaminar normal stress normalized by the mean axial stress
for each of the coupon thicknesses tested is presented in
Figure 4.
107-10
0 (Degrees)
Figure 1: Variations ofDelamination Moment Coefficient (DMC) in
[(±e)n / 90]s Laminates
DMC (10 in)
Figure 2: 6 Ply Laminate, [±40/90]s
Figure 4: Interlaminar Normal Stress Distribution along Midplane of
((±40)n / 90]s Laminates
IV. RECOMMENDATIONS
4.1 Theoretical Studies:
The development to date of extensions to the stress based
theory clearly shows that this approach can be a significant
improvement over currently available alternatives for
describing the stress field in laminated composites subjected
to time dependent loading.
The work presently being conducted should be continued in
order to include;
1) the development of a numerical procedure (fin_te
elements) capable of solving practical problems;
2) continued extension of the method to accommodate a
wider range of problems including material failure and post
impact behavior.
4.2 Experimental Studies;
The numerical studies have shown that an evaluation of a
composite laminate's tendency to delaminate may be made by
identifying a proposed material property, the delaminating
moment coefficient (DMC) . By designing laminates with the
maximum DMC, different composite systems can be made to
delaminate without simultaneously undergoing complicating
additional failure modes.
107-16
Proposed foliow-on studies should include:
1) an experimental program designed to verify the utility
of the DMC in predicting the failure mode. For the graphite
/epoxy system we studied in this effort, we suggest specimens
of several thicknesses each with off axis plies oriented at
±40° be fabricated and tested;
2) additional numerical investigations of different
materials to identify the effects of other system properties
on the calculated value of the DMC.
107-17
REFERENCES
1. Pagano, N.J., "Stress Fields in Composite Laminates,"
International Journal of Solids and Structures . Vol.
14, 1978, pp. 385-400.
2. Sandhu, R.S., "Analytical-Experimental Correlation of
the Behavior of 0°, ±45®, 90® Family of AS/3501-5
Graphite Epoxy Composite Laminates under Uniaxial
Tensile Loading," AFFDL-TR-79-3064 . Wright-Patterson
Air Force Base, Ohio, May, 1979.
3. Sandhu, R.S; and Sendeckyj, G.P., "On Oelaminatibn of
Laminates Subjected to Tensile Loading,
"AFWAL-TR-87-3058 . Wright-Patterson Air Force Base,
Ohio, July, 1987.
4. Schoeppner, 6. A., Composite Laminate Stress Fields
During Dynamic Loading. Final Report, USAF-UES Summer
Graduate Student Research Program, September, 1990.
5. Wolfe, W.E., Low,, Velocity Impact of Graphite/Epoxv
Plates. Final Report, USAF-UES Summer Faculty Research
Program, September, 1987.
6. Wolfe, W.E., Damage in Graphite/Epoxv Plates Subjected
to Low Velocity Impact. Final Report, USAF-UES Summer
Faculty Research Program, September, 1988.
7. Wolfe, W.E. and Schoeppner, G.A., Low Velocity Impact
of Composite Materials. Final Report, Grant Nos. S-760-
7MG-102 and S-210-9MG-082, Universal Energy Systems,
Dayton, Ohio, February, 1990.
107-18
1990 USAF-UES SUMMER FACULTY RESEARCH PROGRAM
GRADUATE STUDENT RESEARCH PROGRAM
Sponsored by the
AIR FORCE OFFICE OF SCIENTIFIC RESEARCH
Conducted by the
Universal Energy Systems, Inc.
Prepared by; Lawrence D. Zavodney, Ph.D.
Academic Rank: Assistant Professor
Department and Engineering Mechanics
University: The Ohio State University
Research Location: Wright Research and Development Center
Flight D3rnafflics Laboratory
Structures Division
Structural Dynamics Branch
Sonic and Fatigue Group (FIBGD)
Wright-Patterson Air Force Base
WPAFB, Ohio 45433-6553
USAF Researcher: Kenneth Wentz
Date: 8 September 1990
Contract No:
F49620-88-C-0053
Experimental Identification of Internally Resonant Nonlinear Systems
Possessing Quadratic Nonlinearity
by
Lawrence D; Zavodney
ABgIBMl
The identification of MOOF nonlinear systems possessing internal
resonance is discussed and possible solution strategies are proposed.
It is shown that it is possible for nonlinear coupling between two
internally resonant modes to go undetected during a modal analysis using
broad'band random*excitati6n. If this type of nonlinearity is not
identified, it is possible that the response to harmonic excitation may
be many times larger than that predicted by the random response. Due to
the combined presence of quadratic coupling and an internal resonance,
it is possible for subharmonic and Hopf bifurcations, combination
resonances, and subharmonic resonances to occur. Nonlinear coupling
terms can also cause excited modes to become saturated. Chaotic
responses were observed and documented. In this report, the results of
experiments conducted at WRDC using conventional and state-of-the-art*
means for system identification are summarized.
108-2
ACKNOWLEDGMENTS
This work was sponsored by the Air Force Systems Command and the
Air Force Office of Scientific Research. Universal Energy Systems, Inc.
administered the Summer Faculty Research Program.
This work would not have been possible without the help of several
key individuals. Dr. Joe Hollkamp provided the technical expertise with
the signal analyzer in collecting and processing the data. The TSSI
technicians who prepared the model did a professional job; they are Mr.
Richard Kleismit and Mr. Paul Reichert. Assistance with the
Instrxunentation and the test facility (shaker) was provided by Mr. Alex
LeDonne and Mr. John Self.
One of the benefits of the Summer Faculty Research Program is the
opportunity to meet and Interact with other researchers at the facility;
to this end the summer experience has an added dimension. My technical
sponsor was Mr. Kenneth Wentz; he coordinated the laboratory facilities
for the experimental phase of this work. The Sonic and Fatigue Group is
managed by Mr. Ralph Shimovetz; the Structural Dynamics Branch Chief is
managed by Mr. Jerome Pearson.
108-3
1. INTRODUCTION
The identification of nonlinear systems has been receiving increasing
attention because most real structures have some degree of nonlinear
behavior associated with them. Problems occur with traditional modal
testing methods because most commercial analyzers run software that assumes
linear behavior, resulting in a best>flt linear model. In general, a
linear model is the least precise because it can always be improved with
nonlinear terms to accomplish three things: 1) to improve the accuracy of
Che model, 2) Co extend the range of useful results by permitting different
amplitude ranges, and 3) to account for behavior unique to nonlinear
systems that has no counterpart in linear theory. This third category is
significant because many nonlinear identification techniques proposed today
are based upon measuring deviation from linear behavior, hence there is an
underlying link to linear behavior from the outset. These methods will be
innacuraCe for nonlinear systems exhibiting uniquely nonlinear behavior.
For work on the ident^icatlon of systems exhibiting uniquely nonlinear
behavior, see Nayfeh^, Zavodney^, and Zavodney and Shihada^.
The significance of an internal resonance is that, when it is
accompanied with a quadratic coupling term, it causes a parametric
excitation to one mode when Che other mode is externally excited (by
driving the structure near resonance) . Since the parametric excitation to
one mode is caused by another mode in the structure, it is called an
autoparametric resonance. This nonlinear resonance can cause subharmonic
bifurcations, the saturation phenomenon, a subcritical instability, or
amplitude- and phase-modulated motions (Hopf bifurcation) . These behaviors
are uniquely nonlinear phenomena; hence, one would not anticipate that
linear based modal analysis techniques would give meaningful results. For
a complete discussion of parametric resonances in mechanical systems, the
reader is referred to Zavodney*.
1084
Quadratic coupling terms can appear in governing equations whenever
there is a quadratic type material nonlinearity, geometrical asymmetry (the
test model used in this report), or a symmetrical structure that is bent,
buckled (beams and plates with in-plaoe edge loads) , or curved.
In this report, conventional vibration analysis techniques are
performed on a "simple" two-degree-of-?reedom (2D0F) lumped-mass model that
can be tuned to possess an internal resonance ( fz - 2fi) . It is shown that
randomly excited nonlinear structures can give seemingly accurate (linear)
results but err significantly in predicting the response to a stationary
harmonic excitation.
2 . TEST MODEL
The experiments were performed on the structure shown in Figure 1;
the length of the lower beam was five inches, its width was 0.5 inches, and
its thickness was 0.030 inch. This structure can model a wing carrying a
store or a space structure with an attachment such as a photovoltaic array.
This structure was chosen for the following reasons:
1) the asymmetrical geometry causes the structure to have nonlinearly
coupled modes, giving rise to quadratic coupling terms,
2) the ease of tuning the structure (by adjusting the length of the
lower beam and the position of the second mass on the upright beam)
to change the natural frequencies of the modes,
3) the possibility of seeing the nonlinear responses without the aid of
a strobe light or dependance upon instrumentation because the
amplitude of motion is lar^e and the mtural frequencies are low.
Two different configurations of the structure vere considered; they were
achieved by changing the location of the mass on the vertical beam. The
two different configurations are:
1) the detuned case where the natural frequencies were 7.69 Hz. and
18.46 Hz. and
2) the tuned case where the frt’.uencies were 5.20 Hz. and 10.60 Hz.
J08-5
Tuning in this case refers to the natural frequencies being conunensurable ,
i.e., in nearly a 2:1 ratio. This structure, subject to harmonic
excitation, has been analyzed by several researchers including Haddow ££
al^. Nay f eh and Zavodney^, and Nayfeh and Balachandran^. The mathematical
models developed show that the structure is dominated by quadratic
nonlinearity that couples the modes. In the absence of an internal
resonance, the structure can be analyzed as a linear system using
conventional modal analysis techniques. However, when the system natural
frequencies are commensurable, the system response is highly nonlinear and
complicated; the two modes are coupled together even though the first 'two
natural frequencies are well separated.
The response of the structure was measured from strain gages attached
near the base of each beam. All displacements (and amplitudes) in the
figures correspond to the displacement wi of mass mi. For the amplifier
gains chosen, one volt corresponded to one inch peak>to-peak displacement.
Additionally, strain rosettes were cemented to the beams to measure shear
strain to Indicate if torsional motion was occurring. During the
experiments, no out-of-plane motion was observed, verifying that even if
the torsional mode was close to or coupled to one of the in-plane modes, it
would not affect the dynamic response or be excited through any of the
nonlinear coupling terms. The structure was clamped to a table excited by
a 12,000-pound force electrodynamic shaker; the data was collected by a
Zonic 7000 analyzer.
3 . RANDOM EXCITATION
The most common modal analysis excitation method uses random noise.
This method excites all of the structural modes simultaneously and hence
reduces the test time. If the analyst suspects that his structure is
nonlinear, he will often excite the structure twice: once with a low-level
excitation and once with a high-level excitation. This dual level
108-6
excitation is often used as an indicator of nonlinearity because linear
systems have two characteristics: the response amplitude is proportional
to the excitation amplitude (the principle of proportionality) and the
response to simultaneous multiple inputs is the sum of the system responses
to the Individual inputs (the principle of superposition). The principle
of proportionality is verified by testing at two different levels of
excitation; if the transfer function does not change with Increasing levels
of excitation, then the system is said to be proportional. Furthermore, if
the system is superposltional , then there would be no difference in the
Frequency-Response Function (FRF) obtained with a random excitation
compared to a sinusoidal excitation.
The procedure to check for proportionality was applied to the
structure. The response of the tuned and detuned structure to low-level
and to high-level random excitation had very similar characteristics; the
response of the tuned structure is shown in Figure 2. There is a very
small change in the FRF, but not the type that is often encountered- -a
shift in and a broadening of the resonant peak when the excitation level is
increased. In this case, there were neither frequency shifts nor
broadening effects, so both the tuned and detuned structures "passed” the
linearity test.
4. HARMONIC EXCITATION
Harmonic excitation methods require an exciter to impart harmonic
forces or displacements to the structure: ah electrodynamic shaker is often
used for this purpose. An oscillator is used to drive the shaker's power
amplifier. In the experiments that follow, harmonic excitation at both
resonances, at a frequency near the sum of both resonances, and at a
frequency near twice the second resonance were performed.
108-7
4.1 Detuned Structure
The frequency response of the detuned structure to a stationary base-
displacement harmonic excitation near the first and second resonance is
shown in Figures 3 and 4. These figures were obtained by sweeping the
frequency up and then down through resonance while keeping the table
acceleration constant. There is no significance in the different
excitation levels for the modes. A 50 milli-g excitation at 5 Hertz
corresponds to a peak- to -peak displacement of about 0.04 inch,
approximately 0.8% ‘of the length of the horizontal beam. Hence, we are
dealing with extremely low-level excitation. The amplitude-response curves
to an excitation of the second mode are shown in Figure 5; this curve was
obtained by fixing the excitation frequency and slowly varying the
amplitude up and then down. Ideal linear systems would exhibit a straight
line indicating that the system response was proportional; the slight
curvature is present because the data are from a real structure.
Another method to detect nonlinearity is to harmonically excite the
structure at resonance and look for harmonics of the driving frequency in
the response. If, for example, there is quadratic nonlinearity present,
then resonances at twice the excitation frequency will appear; if there is
cubic nonlinearity present, then resonances at three times the excitation
frequency will appear in the response spectrum. Quadratic nonlinearity was
detected in the response caused by harmonic excitation at resonance; there
was a large harmonic at two times the excitation frequency. The strain
gage mounted on the vertical beam was the most sensitive to this harmonic.
4.2 Tuned Structure
The frequency response of the second mode to a hairmonic excitation is
shown in Figure 6. As the frequency was slowly increased from below, the
second-mode amplitude grew while the first mode remained trivial (because
it was not excited). However, as the driving frequency approached the
108-8
resonance, the second mode coupled to and excited the first, resulting In
larger steady- state amplitudes. As the frequency was Increased further
(still approaching the second resonance) the amplitudes of the modes
actually decreased! Due to the slight detuning of the internal resonance,
there was no jump phenomenon (turning point, saddle -node or flip
bifurcation) on the left side of the resonance region as there was on the
right side. However, there was a Hopf bifurcation that lead to amplltude-
and phase -modulated steady-state responses for a small region of excitation
frequencies ,
The frequency response of the tuned structure at the first resonance
Is shown In Figure 7. It has behavior similar to that of the second
resonance shown In Figure 6.
The amplitude response of a direct excitation to the second mode, as
shown in Figure 8, reveals three distinct types of responses: the "linear"
region, the nonlinear saturated response characterized by a constant-
amplitude second mode and parametrically excited first mode (growing with
the excitation) , and modulated amplitudes and phases . The bounds on the
mudulation are shown as a maximum and minimum value for the amplitude. The
steady-state response in the modulation region can be easily visualized in
a 3-D spectral history plot as shown in Figure 9. The Eigensystem
Realization Algorithm (ERA)® can track the energy exchange that occurs
between modes In the response by using a sliding window. Figure 10 shows
the ERA sliding window results for the modulated response to a harmonic
excitation. Once the nonlinear behavior is detected. It Is clear that a
linear model is incapable of predicting the response for various levels of
excitation.
Although the frequency-response curves (Figures 6 and 7) and the
amplitude-response curve (Figure 8) show the steady-state amplitude
responses, whenever any system parameter, such as the amplitude or
frequency of excitation, is varied slightly such that a bifurcation
108-9
boundary is crossed, the system dues hot immediately jump to the new steady
state as shown on the plots. Instead^ the response grows with each cycle
of motion towards the steady-state amplitude. Figure 8 shows the
theoretical linear solution as dashed lines; if the system is in "steady-
state" below the critical value of excitation causing saturation and the
amplitude of excitation is slowly increased to a value above the critical
value, the system will behave as shown in Figure 11. The second mode will
initially grow to the unstable linear response, but as time increases, the
first mode (which is parametrically excited) begins its exponential growth.
As it grows, it takes energy from the second mode which is clearly seen.
4.3. Combination and Subharmonic Resonances
One of the unique properties of nonlinear systems is that they
possess resonances not found in linear systems. In this case, the
nonlinear coupling between the two modes allows additional resonances to be
present in the structural response to harmonic excitation that are not
excited with broad-band random excitation. In particular, one of the
quadratic coupling terms is responsible for a combination resonance
whenever the excitation frequency is close to the sum of the frequencies of
the two modes; another quadratic term is responsible for a one-half
subharmonic bifurcation which gives rise to a subharmonic resonance. When
the system simultaneously possesses an internal resonance, the modal
interactions observed in the direct excitation of either mode can also be
observed in the indirectly excited responses.
For example, when the structure is excited at a frequency two times
the second natural frequency, the response will consist of the forced
response (particular solution of the governing differential equation at the
driving frequency) and a subharmonic of the excitation frequency, which
happens to coincide with the second resonance. This behavior is shown in
Figure 12(a). It is possible to couple the first mode when an internal
108-10
resonance exists just by changing the excitation frequency or amplitude
slightly, as shown in Figure 12(b). Furthermore, it is also possible for
the coupled modes to modulate as previously seen, as shorn in Figure 12(c).
By zooming in on the resonance region of the spectrum, the subharmonics can
be more easily seen; this is shown in Figure 13.
When the structure is excited by a frequency near the sum of the
first and second resonant frequencies (a combination resonance of the sum
type), the two modes will be simultaneously be excited. For certain ranges
of system parameters, modulated responses can again be observed; Figure 14
shows the spectral history observed when the structure was excited with a
stationary constant-amplitude acceleration. Using the ERA sliding window,
the modal amplitudes are observed to modulate while the particular solution
remains essentially constant, as shown in Figure 15.
5. IMPULSE EXCITATION
Modal analysis using impulse excitation is quite appealing for field
work because it does not require a shaker, power amplifier, or signal
generator. With an Instrumented hammer, a complete modal analysis can be
performed with a two-channel analyzer. If the structure is linear, the
random excitation, impulse excitation, and swept sine excitation will give
identical results. Hence, if the system is nonlinear, one would expect the
different methods to produce different results depending upon the type of
nonlinearity present in the structure.
For this experiment, mi was hit with a small hammer avid the time
history of the response was recorded. Modal interaction was present in the
free response of the tuned structure; however, it would not be obvious to
the untrained analyst. It is not uncommon to observe a beat phenomenon '^n
a free response of a MOOF structure; typically all of the modes are
excited, and as they decay at their damped natural frequencies, they will
combine to form a composite response that may beat. However, unless the
108-11
Individual modal ampHtudes are extracted from the composite response
signal, the interaction would most likely be obscured. A transient
response of the stxMcture to an impulse iA shown in Figure 16; the modal
decomposition (extracted with the aid of the ERA sliding window) is shown
in Figure 17. This Figure shows the energy being exchanged from one mode
to another while both are simultaneously decaying. These results are
similar to those obtained by Zavodney and Pappa^ for the free response of a
similar structure. If there was no internal resonance, each mode would
exponentially decay from its original amplitude.
6. TIME SERIES ANALYSIS
The method of delays was applied to the time series to construct
pseudo-phase planes. This method uses the time series representing the
displacement and constructs the "pseudo" velocity by using the position at
a later time. In this case, a delay of 0.025 seconds was used to construct
the phase plane . The response, of the tuned s,t.'ucture when the excitation
frequency was near the second resonance was analyzed because it had three
distinct responses (as shown in Figure 8) . the phase portraits are shown
in Figure 18; the data show typical scatter about the equilibrium
attractor. Poincare maps were obtained from the phase planes and are shown
in Figure 19. Only a portion of the modulated response is shown in the
phase plane because the complete cycle would completely fill in the bounded
region; the Poincare map is constructed from the entire set of 60,000
points. To construct the Poincare maps, the excitation signal was filtered
with an 8^**- order low-pass ARMA filter before it was used to sample the
phase plane. The modulated response appears to be chaotic.
108-12
7. SUMMARY
1. Random excitation may give results that appear linear, making it
possible for nonlinearity to go undetected. One reason why the random
excitation would tend to obscure the autoparametric resonance is that
the modal coupling is parametric in nature; parametric resonances feed
on themselves and grow exponentially. Consequently, there is
insufficient time for a' parametric resonance to grow when the
excitation is random.
2. Muiti-level random excitation will not reveal certain types of
nonlinearity because the nonlinearity in this case is not associated
with the elastic stiffness; instead the nonlinearity is primarily from
the geometry. Hence, increasing the excitation would not cause
changes in the frequency response because the elastic stiffness is not
strongly dependant upon the amplitude of response.
3. Low-level harmonic excitation can cause linear responses or nonlinear
responses with constant or nonconstant amplitudes that modulate in
time.
4. Subharmonic resonances and combination resonances may exist that could
produce large responses . •
5. Impulse responses may have modulated responses that are embedded and
may not be obvious; they would need to be extracted to be identified.
108-13
6. CONCLUSIONS
1. System identification should include sufficient checks for nonlinear
resonances caused by material properties, geometrical as3rmmetry, and
stiffness nonlinearity, even when the data obtained from low and high
level random excitation do not suggest any nonlinear behavior.
2. System identification should include all possible types of excitation a
structure could possibly be exposed to; this would insure that
xuidetected nonlinear resonances using one type of excitation would not
appear later using a different type of excitation.
3. Slight structural modifications after a design has been qualified could
cause nonlinear resonances to appear. These slight modifications
could cause, for example, ah internal resonance to occur, and provide
the mechanism for the observed nonl:.near behavior documented in this
report.
4. If a structure demonstrates autoparametric coupling in response to a
harmonic excitation, random excitation will effectively prevent the
nonlinear coupling from occuring.
5. An identification technique that is uniquely suited to nonlinear
behavior needs to be implemented for the more difficult structures.
This technique would use the nonlinear resonances and jumps in the
108-14
system response to
determine quaHtatively and quantitatively the best
mathematical model for the structure.
8. REFERENCES
1. Nayfeh A.H. , "Random Motion and Dynamic Response<Faraffletric
Identification of Nonlinear Dynamic Systems," Computers and
Structures. Vol. 20, No. 1, pp. 487-493, 1985,
2. Zavodney L.D. , "Can the Modal Analyst Afford to be ignorant of
Nonlinear Vibration Phenomena," Proceedings of the Fifth
International Modal Analysis Conference. 1987,
3. Zavodney L.D. and Shihada S.M. , "The Identification of Nonlinearity
in Structural Systems: Theory and Experiment," The Third Conference
on Nonlinear Vibrations, Stability, and Dynamics of Structures and
Mechanisms, Blacksburg Virginia, June 1990.
4. Zavodney L.D., "A Theoretical and Experimental Investigation of
Parametrically Excited Nonlinear Mechanical Systems," Ph D.
Dissertation, Virginia Polytechnic Institute and State University,
1987.
5. Haddov A.G., Barr A.D.S., and Mook D.T., "Theoretical and
Experimental Study of Modal Interaction in a Two-Degree-of-Freedom
Structure," Journal of Sound and Vibration. Vol. .97, pp. 451-473,
1984.
6. Nayfeh A.H. and Zavodney L.D., "Experimental Observation of
Amplitude- and Phase- Modulated Response of Two Internally Coupled
Oscillators to a Harmonic Excitation," Journal of Applied Mechanics.
Vol. 55, pp. 706-710, 1988.
7. Nayfeh A.H. and Balachandran B. , "Experimental Investigation of
Resonantly Forced Oscillations of a Two -Degree -Of -Freedom Structure,"
International Journal of Nonlinear Mechanics. Vol. 25, pp. 199-209,
1990.
8. Juang J-N. and Pappa R.S., "An Eigensystem Realization Algorithm for
Modal Parameter Identification and Model Reduction, Journal of
Guidance. Control and Dynamics. Vol 8, pp. 620-627, 1985.
9. Zavodney L.D. and Pappa R.S., Unpublished ERA results obtained during
a NASA Langley/ASEE Summer Faculty Research Program in 1988.
108-15
Figure 1, Two-degree-of-freedom model pos-
aesalng quadratic nonlinearity; by
adjusting the lengtha of the
beama, the atructure can be tuned
for a 2:1 Internal reaonance.
The Inaeta ahow the linear mode
ahapea.
0 10 20 30 40 30 60
Frequency (Hz)
Figure 2. Frequency-Response Function of the
tuned structure for random excita¬
tion at two different levels. The
FRF of the detuned structure at
two different Input levels also
correspond to each other.
T 7.2 7.4 7» 7,8 8
ExeltaHon Fraqueney (Hz)
Figure 3. Frequency-Reeponae Function of the
detuned structure for stationary
sinusoidal excitation at the first
resonance. The excitation level
Is 50 nilli-g's, fi ° 7.69 Hz, fo
- 18.52 Hz.
n
t
t
p
0
n
*
V
/
s
18 182 184 188 188 18
Excitation Frequency (Hz)
Figure 4. Frequency-Response Function of the
detuned structure for stationary
sinusiodal excitation at the
second resonance. The excitation
level is 21.2 raili-g's, f. = 7.69
Hz, f, = 18.52 Hz..
0 20 40 $0 so 10O
Excitation Ampiitude (miiii-g's)
Figure 5. Amplitude response of the detuned
structure for harmonic excitation
of the second mode showing typical
linear behavior. An ideal struc¬
ture's response would be perfectly
straight .
108-16
20
to lO.t tO.2 103 tO.4 10.S too 10.7 10« 10* It
Excitation Fraquancy (Hz)
03 r
MMMt O MMMt
oa^ - - - Oooo-
OoOOqOOoO oOOO
o
✓o
0.1 1-. - J,... O.. ~ .. — .....
o . • •
o •
0 I 10 IS
Excitation Amplitud* (niiiii.o's)
10
Figure 6. Frequency response of the tuned
structure for a harmonic excita¬
tion near the second resonance.
Note that the first mode is non-
linearly coupled to the second
mode; for some frequencies there
is no steady-state response. The
excitation level is 21.2 milli-
g's, f, - 5.2 Hz., f, - 10.6 Hz.
Figure 8. Amplitude response of the tuned
structure for harmonic excitation
of the second mode. Note that for
low levels of excitation the re¬
sponse is similar to the linear
detuned case; at a critictil level
of excitation the second mode is
saturated and the first mode ap¬
pears in the response. This be¬
havior is similar to parametric
resonances of nonlinear systems.
At the second critical level of
excitation a Hopf bifurcation
occurs and the steady-state re¬
sponse consists of .modulated
amplitudes and phases.
Figure 7. Frequency response of the tuned
structure for a harmonic excita¬
tion near the first resonance.
Note that the first mode is non-
llnearly coupled to the fcond
mode; for some frequencies there
is no steady-state response. The
excitation level is 50 milli-g's,
fj = 5.2 Hz., fj = 10.6 Hz.
•••• 11. • tta.
Figure 9. Waterfall plot showing the spec¬
tral history of the modulated re¬
sponse of the structure to a har¬
monic excitation at the second
resonant frequency.
108-17
CO
Figure 18. Pseudo phase portraits of a direct
excitation of the second mode (of
the tuned structure) constructed
from the displacement using the
method of delays: (a) harmonic
linear response, (b) subharmonic
bifurcation, and (c) chaotic modu¬
lated response. See Figures 6
through 10.
Figure 19. Poincare maps of the phase por¬
traits shown in Figure 18: (a)
single point showing periodic
response with the excitation, (b)
bifurcated response showing a
period 2 response, and (c> a
strange attractor showing chaotic
behavior.
108-20
1990 USAF-UES SUMMER FAGLTY RESEARCH PROGRAM/
GRADUATE STUDENT RESEARCH PROGRAM
Sponsored by the
AIR FORCE OFFICE OF SCIENTIFIC RESEARCH
Conducted by the
Universal Energy Systems, Inc.
FINAL REPORT
Ihfi,. Jnr^u Laser . Deposition of . High Tc
Superconducting . Thin ..Film.
Prepared by:
Academic Rank:
Department and
University:
Research Location:
USAF Researcher:
Date:
Contract No.:
Donald D. W. Chung, Ph.D
Associate Professor
Materials Science
San Jose State University
USAFWRDC/MLPO
Patrick M. Hemenger
27 Aug 90
F49620-88-C-0053
Ihin-FUm
by
Donald D. W, Chung
ABSTRACT
ArF excimer laser ablation of an YBa2Cu307-x target pellet in 100
mTorr of O2 ambient was used to deposit thin superconducting films
on SrTiOs and MgO substrates at 650 - 780 OC. The as-deposited 0.6
- 0.9 |im thick films were superconducting, without further high-
temperature annealing. Cooled to ambient temperature in -situ for
1.5 hours in flowing oxygen gas, the films showed complete
diamagnetism and zero resistance up to 89 K with a critical current
density of 5x1 05 A/cm^ in zero magnetic field at 81 K. Low angle X-
ray diffraction analysis showed that all the films were highly
oriented with the C-axis perpendicular to their surface. Smooth
surface morphology was observed in all films.
109-2
Acknowledgements
I wish to thank the Air Force Systems Command and the Air
Force Office of Scientific Research for sponsorship of this research.
Universal Energy Systems must be mentioned for their concern and
help to me all administrative and directional aspects of this program.
My experience was rewarding and enriching because of many
different influences. Dr. Patrick Hemenger provided me with
support, encouragement, and a truly enjoyable working atmosphere.
The help of Tim Peterson was invaluable overcoming many technical
roadblocks. Dr. Iman Maartense’s interest in every phase of this
project truly served as source of stimulation. The encouragement
and help of Dr. Terry Murray were deeply appreciated.
I. Introduction:
Considerable progress has been made in the field of thin film
preparation of high critical temperature superconducting materials
both for fundamental research and possible application in electronic
devices. Such film preparation is being attempted by many film
deposition techniques. Among them, one promising method has been
by mean of laser ablation of a bulk target superconducting materials
in vacuum or under the low oxygen pressure to deposit the film.
Since the low temperature processing will be essential for making
sophisticated electronic devices containing the superconducting films
the laser deposition method for such film preparation are especially
attractive in the potential for producing films with required heat
treatment in oxygen at only modest temperatures. Among the
various types of laser processes short pulsed laser deposition has
provided excellent epitaxial films with shallow resistance transition
widths and high critical currents. Various researchers have obtained
superconducting films without high temperature post-annealing
using a background O2 ambient in the vacuum chamber.
The elimination of the post-annealing process is extremely
important in terms of growing multiple layers of different materials
for device applications. Recently the in-situ processing of
superconducting films at substrate temperatures of 650 - TSO^C has
been reported by a variety of techniques. The introduction of
oxygen jet into the laser, ablated plasma was found to be a key for
in-situ processing of superconducting thin films at bSO^C. Further
1094
reduction in deposition temperature have been reported by the
incorporation of an oxygen plasma. So far* epitaxial film growth of
YBa2Cu307-x on SrTiOs has yielded films with the highest current
densities, the sharpest resistance transitions, and in general, the best
overall properties.
Superconducting thin film processes and control programs are
particularly interested in space application for detectors and sensor
devices in Materials Laboratory of MLPO of the USAF. A number of
sensor devices have been demonstrated with high Tc
superconducting thin films; namely, superconducting quantum
interference devices (squids), tunnel junction, and a fast nonlinear
switch for noise discrimination in digital circuits. Special attention is
directed to the optical and infrared detection using these new
materials.
My research interest have been in the area of application of
variety of thin film deposition techniques in the investigation of
electrical, magnetic, and structural properties as function of
processing parameters. My work on the deposition of YBa2Cu307-x
superconducting film by RF and DC magnetron sputtering in which
the principle of deposition is complementary to the laser ablation
deposition, which is in principle, the same physical vapor deposition,
and characterization of the film in terms of electrical, magnetic and
structural properties contributed to my assignment to the
superconductor group; Materials Laboratory, MLPO division of the
USAF.
109-5
n. OBJECTIVES OF THE RESEARCH EFFORT
The thin film processes of high Tc superconducting YBa2Cu307-x
compound can be divided into what require a high temperature
annealing subsequent to deposition, and what require no further
annealing, or low temperature annealing in oxygen to achieve
optimum superconductivity. Low temperature deposition techniques
which do hot require further annealing are, by far, the most
promising technique for producing films for applications because a
better film morphology is usually obtained. The pulsed laser
deposition is on of the most successful methods of achieving low
temperature films in which it is capable of processing very smooth
films with high transition temperature and critical current densities.
However, further investigation is called for in this process in order
for these to be a definitive evaluation of this technique with respect
to applicability of the grown film properties as practical device
materials.
During my 1989 SFRP several film processing data was
generated and collected. The important deposition conditions were
determined since the film processing data could not be analyzed in
time. Research in this project resumed at San Jose State during
January of this year under the sponsorship of an AFOSR Mini Grant.
The significant processing conditions had been established prior to
the 1990 SFRP at Wright-Patterson Air Force Base. Preparation of
high Tc superconducting thin films are now routinely performed.
109-6
My assignment as a participant in the 1990 Summer Faculty
Research Program (SFRP) was to continue to study arid determine the
process and the properties for depositing high quality temperature
superconducting films and assess parameters of the laser deposition
conditions for optimum properties which would be required for
suitable device structure applications such as electronic detectors
and sensor materials.
III.
High quality thin YBa2Cu307-x film has been produced by laser
ablation technique. In many cases an amorphous film is deposited
and a post annealing step up to 900®C is necessary to form the
crystalline superconducting phase. However, a good orientation of
these films, a minimization of grain boundaries, and the prevention
of inter diffusion are important points in order to achieve high
critical current densities. All of the above is easier to obtain with an
in-situ growth process, as has recently been demonstrated by several
research groups.
For our experiments, we used a ArF excimer laser with 193 nm
wavelength, 20 Hz, and 80 mJ per pulse. The laser beam was focused
on a rotating, sintered YBa2Cu307-x pellet in a vacuum chamber.
The target material which was evaporated perpendicularly to the
substrate surface was then deposited on the substrate at a distance
of about 60 mm. The substrate was heated up to 850OC for cleaning
and then cooled to the deposition temperature. Oxygen can be added
109-7
both the deposition and cooling processes. The chamber was
evacuated by a turbomolecular pump; the base pressure was less
than 10'^ Torr. Single crystalline SrTiOs and MgO with < 100 >
orientation were used as substrate materials. The polished
substrates were cleaned in an ultrasonic bath in trichloroethylene,
acetone, methanol, and de>ionized water, in that order and heat
treated at SSO^C for half an hour in-situ before the deposition. The
deposition was carried out in about 30 minutes with a typical
deposition rate of 4 A/sec, and the resulted in film thicknesses
ranged from O.b^im to 0,9pm. Immediately after the deposition, pure
oxygen was introduced into the deposition chamber. The films were
then cooled down in flowing oxygen gas to ambient temperature in
1.5 hours.
X-ray diffraction analysis was performed with a slow scan
diffractometer. The films were further examined by four-probe
resistivity and A.C. susceptibility measurements. For the
measurements of resistivity, the films were patterned by sputtering
bridge paths 5mm length and 100 pm width. Four silver contacts
were evaporated onto the sample and bonded with gold wires.
To determine the desired film processes, we first investigated
the film properties as a function of several deposition parameters
such as the substrate temperature, oxygen pressure, and target-
substrate distance. The film properties were then evaluated by A.C.
magnetic susceptibility measurement and X-ray diffraction analysis.
The in-situ film growth behavior was depending sensitively on the
109-8
substrate temperature and the oxygen pressure. Films deposited at a
substrate temperature lower than 600° C were semi-crystalline and
had to be annealed to 900° C to form the desired crystal structure.
At a substrate temperature higher than 650° C the film had grown
in-situ crystalline. For the epitaxial growth with the C axis
perpendicular to the substrate plane, oxygen pressure of 100 mTorr
and a substrate temperature between 730 - TSO^C were optimal.
The X-ray diffraction pattern indeed showed a highly < 001 >
oriented film grow expitaxially on both SrTiOs and MgO.
The quality of the C-axis orientation was also documented by
observing the (OOn) peaks from the x-ray diffraction pattern. It has
been reported' in the literature that the occupation of the oxygen
sites in the lattice is strongly correlated with the C-axis parameter
and the critical temperature. Films quenched after the deposition
showed superconductivity but with low Tc components. Samples
cooled slowly, generally showed high quality superconductivity.
Although most of the films were exposed to the substrate
temperature above TSO^C, little interdiffusion between the substrate
and the film had taken place. For the majority of the in-situ film the
resistivity behavior was metallic. We have observed critical
transport current density values Jc above 10 A/cm^ at 81 K.
In conclusion, we have succeeded in preparing epitaxial
YBa2Cu307,x films with high critical current densities by laser
ablation in an easily reproducible one step process. No additional
oxygen plasma source was needed for the in-situ crystalline growth.
109-9
IV. RECOMMENDATIONS:
The in-situ deposition of high critical temperature
superconducting films ( YBa2Cu307-x ) on SrTi03 < 100 > substrate
by an ArF excimer laser have been successfully accomplished
without additional system modification. However, the evaluation of
data showed a room for improvement in the future deposition
condition by implementing the existing system modification in the
following areas:
a. The actual substrate (sample) temperature measurement
must be established by installing a new substrate holder stage
(including heat source) perhaps with a silver block which may be
rotated during deposition.
b. The laser power and pulse rate during deposition must be
controlled and optimized.
c. The experimental condition for low temperature deposition
must be established for device applications.
d. The oxygen content of the film after the deposition and
cooling process must be monitored in junction with the film property
change.
e. The high dielectric constant and loss tangent of SrTiOs
substrate would limit its practical utility, particularly in high-
frequency microelectronic applications. Some other types of
substrate such as NdGaO, LaGaO, or LaAlO which have good dielectric
constants at high frequencies must be established for laser
deposition conditions.
109-10
REFERENCES
Bauerle, D., Laser-Induced Formation and Surface Processing of High-
Temperature Superconductors, Appl. Phvs.. 1989, Vol. 48, pp. 527-
542,
Koren, G., Gupta, A., and Baseman, R.J., Role of Atomic Oxygen in Low-
Temperature Growth of YBa2Cu307-x Thin Films by Laser Ablation
Deposition, Appl. Phvs. Lett.. 1989, Vol. 54(19), pp. 1920-1922.
Mogro-Campero, A., et, al., Epitaxial Growth and Critical Current
Density of Thin Films of YBa2Cu307.x on LaA103 Substrate, Appl.
Phvs. Lett.. 1989, Vol. 54(26), pp. 2719-2721.
Singh, R.K., et. al., In-situ Processing of Epitaxial Y-Ba-Cu-0 High Tc
Superconducting Films on (100) SrTiO and (100) YS-ZrO Substrates at
500-650OC. Appl. Phvs. Lett.. 1989, Vol. 54(22), pp. 2271-2273.
Witanachchi, S., Kwok, H.S., Wang, X.W., and Shaw, D.T., Deposition of
Superconducting Y-Ba-Cu-0 Films at 400OC without Post Annealing,
Appl. Phvs. Lett.. 1988, Vol. 53, pp.234-236.
109-11
1990 USAF-UES SUMMER FACULTY RESEARCH PROGRAM/
GRADUATE STUDENT RESEARCH PROGRAM
Sponsored by the
AIR FORCE OFRCER OF SCIENTTFIC RESEARCH
Conducted by the
Universal Energy Systems, Inc.
EINAL RETORT
AMI CALCULATIONS ON RIGID ROD POLYMER MODEL COMPOUNDS
Prepared by:
John W. Connolly
Academic Rank:
Professor
Department and
Chemistry
University:
University of Missouri*Kansas City
Research Location:
WRDC/MLBP
Wright Patterson Air Force Base, Ohio 45433-6533
USAF Researcher
Douglas S. Dudis
Date:
1 1 Aug. 90
Contract.
F49620-88-C-0053
by
John W. Connolly
ABSTRACT
Using AMI semi-empirical Molecular Orbital calculations, conformational energies were
obtained for structims designed to model the rigid rod polymers, poly(p-
phenylenebenzobisoxazole), PBO, poly(p-phenylenebenzobisimidazole), PBI, and poly(p-
phenylenebenzobisthiazole), PBT, including examples in which the phenylene ^up is
mono and dimethlylated. Minimum energy torsional angles and barriers to rotation can be
understood in terms of steric factors and disruption of pi-electron delocalization. The model
system used shows that when adjacent segments of the polymer chain are mutually
perpendicular, the bturier to rotation is less than the thermally avalable energy at SOOK.
110-2
Acknowledgements
I wish to thank the Air Force Systems Command and the Air Force Office
of Scientific Research for sponsorship of this research. The Assistance of
the personnel of Universal Energy Systems, especially Mr. Danishek, with
the every-day aspects of this project is appreciated.
The atmosphere in the Computational Chemistry Group of the Polymer
Branch at the Materials Laboratory was made enjoyable by the enthusiasm
and good cheer of Doug Dudis. The patience and help of the other members
of the group, especially Todd Yeates and Jacque Henes is gratefully
acknowledged.
110-3
I.
mTRQDUCIIQH:
The rigid ixxi polymers poly(p-phenyIenebehzobisoxazole), PBO, and its sulfur analog,
poly(p-phenylenebenzobisthiazole), PBT, are the focus of the U. S. Air Force's "Ordered
Polymers" Program, which has been established to develop low-weight, high-performance
materials for military and aerospace applicadons. Thesematerials have been found to
exhibit exceptional specific strength and , thermooxidative stability and environmental
resistance when made into rilms and fibers (1).
The Computational Chemistry Group of the Polymer Branch of the Materials Laboratory at
the Wright Research and Development Corporation is investigating the properties of PBO
and PBT model compounds, primarily through the use of semi-empirical molecular orbital
calculations. A wide range of polymer properties, including electronic, optical and
mechanical, can be modelled using the appropriate o^culational technique.
My research experience in the use of mainframe computers both for NMR spectra
simulation and for semi-empirical molecular orbital calculations on organometallic
compounds contributed to my assignment to the Computational Chemistry Group.
II. OBJECTIVES QF.THE RESEARCH EFEQRT:
There have been two recent AMI molecular orbital (MO) calculations on the conformation
energy of PBO and PBT model compounds, (2,3).The structures used in these calculations
are shown in Figure 1. In both cases rotational barriers for PBO of 5.0 KcaVMol and 1.7
110-4
KciJ/Mol for PBT were reported. However a molecular modelling by Fanner, as yet
unpublished, indicated greater polymer chain flexibility than is compatible with the above
results. Consequently, it was decided to calculate conformation^ energies for PBO and
PBT model compounds in which rotation about adjacent caibcm-caibon bonds in the
polymer backbone need not have the same barrier height.
Figure 1. Model Compound Structures used in Previous AMI
Rigid Rod Studies
110-5
m
The calculations done here were penoimed using the AMI (4) Kim empirical method as
implemented by the AMPAC 2.01 series of programs, which is available from ^e Quantum
Chemistry Exchange Program, Indiana University, Bloomington, IN 47405 as QCPE
#506. The newly introduced sulfur parameters were used in the calculations (5). The model
compound used in our calculations is shown in Figure 2. The phenyl group was rotated
paddle-v/heel fashion about the carbon-carbon bonds indicated in the structure and heats of
formation were calculated at every 10® rotation.
Figure 2. Rigid Rod Model Compound Structure Used in This Study
Table 1 shows a complete listing of structural types which were investigated; specific
structures investigated have X=0, X=NH, and X=S. Thus, in addition to the all-plan^
structures (I, HI, V, VII), we have investigated conformations in which the two
heterocyclic groups in the structure are mutually perpendicular. Also we have investigated
the effect of phenyl substitution on rotational barrier height (Hl-VIII).
110-6
TABLE 1
Structural Types on which AMI Calculations were done
SYMBOL
STRUCTURE
1
C0“O“C0
II
III
CO“^)“CO
IV
V
CO-^-CO
VI
VII
VIII
110-7
Figure 3 shows a typical result of the calculation described here, the barrier height is the
difference between zero and the curve maximum on the Y axis. This figure shows the
significant result that when the two ends of the structure are perpendicular the barrier height
is less than 1 Kcal/Mol (cuh^e C).
p-BIS(BENZOXAZOLE)BENZENE
-20 0 20 40 60 80 100
TORSION ANGLE
Figure 3. Conformational Energy Curves for PBO Model Compounds.
110-8
Table 2 shows a compete summary of all the calculations done during the summer 1990
period. In summary, we find that in all the model systems investigated here the calculated
TABLE 2
TORSION ANGLE
TORSION ANGLE
BARRIER
(DEG)
(DEG)
(Kcal/Mol)
STRUCTURAL MIN ENERGY
MAX ENERGY
TYPE
X=0 X»NH X«S
X-0 X»NH X»S
X=0 X=NH
X=S
I
0
30
10
90
90
90
5.2
2.7
2.2
II
45
45
45
0,90
0,90
0,90
0.5
1.3,
0.5
in
6
140
30
90
180
90
4
3.1
1.3
IV
20
150
60
90
0
0
1.7
4
1.6
V
40
50
70
90
0
0
1.7
7.1
1.7
VI
20
50
30
%
0
0
4.2
8.5
3.7
vn
80
50
90
90
0
0
2.5
4.7
0.2
vra
140
130
45
0
180
0
1
3.1
1.
In summary, we find that in all the model systems investigated the calculated barrier height
is drastically reduced in the structure where the two heterocyclic groups are mutually
perpendicular. Since this represents a physically achievable conformation in the polymer
chain, it suggests that the rigid rod polymers models in this study may be more flexible
than had previously been thought
The loss of resonance energy as the phenyl group rotates is a major factor in the
magnitude of the rotational barrier in all the systems we investigated. Substitution on the
central phenyl group causes steric interactions to become a factor as well. In nearly every
case we examine methylation c ' the central phenyl group decreases the rotational barrier
since it destabilizes the (fi torsion angle structure. In the case of structtire type VI this
110-9
destabilization becomes so great that the 0° structure is the highest energy conformation
the rotational barrier is increased.
IV.
Bond Order Output from AMI calculations yields informadon about bond order and
atomic charge in addition to the heats of formation described above. The change in bond
order between Cl and C7 (Fig. 2), while quite small, is an indication of the change in pi-
electron delocalization between the arylene ring and the rest of the molecule. We found that
the C1-C7 bond order decreased monotonically by a total of about 3% as the torsion angle
went from 0° to 90°* The calculated bond order is evidently not sensitive to steric
interaction, but it has the advantage that it isolates the electronic aspect of the barrier height
in every case.
110-10
Figure 4 shows the excellent corielation between bcMid order and barrier height for
the o^methylated PBO species (structures in and IV in Table 2).
TORSION ANGLE
Figure 4. Bond Order<Torsion Energy Correlation for PBO Model
Compounds.
110-11
On the the other hand, Figure 5 shows that the bond order-rotational barrier correlation is
not simple for the corresponding irhidazole species. Since the bond order-torsion angle
curve is the same for all the species examined here it is clear that non electron-delocalization
factors cause the more complex rotational barrier curves seen in most of the species studied
here.
Figure 5. Bond Order-Rotation Barrier Correlation for FBI Model
Compounds.
110-12
V.
Atomic Charge: The AMI results obtained here indicate that minor charge redistribudoh
occurs as the torsion angle changes. Charge flows away from the imine nitrogen as
coplanarity is lost and charge flows away from the oxazble oxygen as cqplanariQr is lost in;
all cases. The charge on the amine nitrogen is unaffected by change in the torsion angle. In
the thiazole series charge flows away from the sulfur as coplanaiity is lost except for the
sulfur adjacent to the two methyl groups in the oitho-dimethyl species, where charge flows
toward sulfur as the torsion angle increases. As charge flows away from the heterocyclic
atoms, Cl becomes more negative which is consistent the decrease in electron
delocalization as coplantuity is lost Overall atomic charge changes were less than .03
charge units in all cases. Charge changes do not correlate with barrier height in any simple
way which is illustrated in Figure 6.
110-13
Figure 6. Atomic Charge*Rotational Barrier Correlation for FBI Model
Compounds.
VI.
Structural Comparison: It has already been demonstrated that the energy-minimized
AMI structures are very similar to those determined experimentally for PBO and PBT
model compounds (6). The major discrepancies are that in the AMI structures the oxazole
oxygen atoms and the thiazole sulfur atoms appear to be closer to sp^ hybridization while in
the experimental structures these atoms are closer to sp^ hybridization.
110-14
The crystal structure of l,4-bis(2-benzoxaz6lyl)-2,S-bis(2*benziinidazoIyl)benKne, Figure
7, has been determined (7). While the steiic inactions between adjacent heterocyclic rings
in this ctnnpound are significant, the oxazole ring is only 5° from coplanarity with the
central phenyl group while the torsion angle of the imidazolyl ring is 57°. This is
consistent with our findings as well as those of Farmer (3) regarding FBI model
compounds. .In Table 5 we compare some experimental and calculated bond distances and
angles. We chose the 0° torsion angle PBO model and the 60° torsion angle FBI calculated
species for our comparison.
110-15
TABLES
Selected Values of Observed wd AMI Calculated Bond Lengths and Bond Angles for 1,4-
Bis(2-benzoxazolyl)-2,5-bis(2-ben2iniidazolyl)bchzene and PBO and PBI Model
Compounds
Bond Lengths
Bond*
Exp. Calc.
Bond
Exp.
Calc.
C1-C7
1.466 1.453'
C7-N1
1.329
1.332
C7-01
1.347 1.433
C13-N1
L392
1.412
C8-01
1.385 1.396
C2-C14
1.490
1.467
C14-N3
1.393 1.400
Ci4-N2
1.319
1.351
C1-C2 1.412 1.400
Bond Angles
Bond*
Exp.
Calc.
C2-C1-C7
125.4
1-19.9
C1-C7-N1
126.3
113.7
C1-C7.01
119.5
115.7
C7-N1-C13
105.0
104.2
C1-C2-C14
125.7
119.2
C2-C14-N2
123.8
126.0
C14-N3-a0
104.3
106.5
C14-N2-C15
105.7
105.4
*The numbering system is as in Figure 7; analogous atoms in Figure 2 were used.
110-16
I
Vn. RECOMMENDATIONS:
The above calcidp.tions have aU been done on neutral, electron-paired species.
Actual PBT fibers are processed at high acid strength (1) and the resulting dope is
subjected to mechanical stress during the foimation polymer films. Electron Spin
Resonance, E3R, measurements made on unstressed PBT fibers show the presence of free
radicals. The ESR signals of the fibers diminished on annealing and increased when the
annealed fiber samples were stretched or crushed. A reasonable speculation is that the
radicals in the unstressed fibers are ion radicals produced during strong acid workup, and
the radicals induced mechanically are neutral r^cals resulting from breaking polymer
backbone (8). The presence of an unpaired electron in the PBT polymer may well ^ect the
stiffness appreciably.
It is possible to do AMI MQ calculations on radical species as well is electron-
paired species. The complications with odd-electron species are that unrestricted Hartree-
Fock (UHF) calculations must be done, which are CPU time intensive, and the energy
minimization process may not converge. We recommend then that a thorough, systematic
series of MO calculations be done cm the cation radical, anion radical, and neutral radical
species derived from the even-electron compounds already reported on here. Our primary
interest is in the effect of unpaired spin density on the barrier to rotation about the
backbone carbon-carbon bond as well as bond order changes within the polymer backbone.
Besides having implications for mechiuiical properties such "defects" are also relevant to
postulated electrical conduction mechanisms in these and similar organic materials.
110-17
REFEI^NCES
1) Evers R..C, Arnold, F. E., Helminiak, T. E., "Articulated All-Para Polymers with 2,6-
Behzobisoxazole, 2,6-Benrobisthiazole, and 2, 6 Benzobisiriiidazole Units in the
Backbone j. Am. Chem. Soc. (1981), 14, 925
2) Welsh W. J. and Yang, Yong AMI MO Calculations of the Cwiformational
Characteristics of the Rodlike Polymers PBO and PBT Macromoleules, (1990), 23,
2410
3) Farmer, B. L., Wiershkke, S, G., and Adams, W. W., Study of the Conformations of
Sdff Chain Rigid Rod and Substituted Rigid-rod Polymers, Ibid, in press
4) Dewar, M. J. S., et al., AMI: A New General Purpose Quantum Mechanical Molecular
Model, J. Am.Chem. Soc., (1985), 107, 3902
5) Dewar, M. J. S. and Yang, Y-C, Inorganic Chemistry, in press
6) Welsh, W. J. and Mark, J. E., "Applications of Quantum Mechanical Techniques to
Rigid Rod Polymers, in "Molecular Level Calculations of the Structure and
Properties of Non Crystalline Polymers", Bicerano, J., Ed. Wiley in press
7) Fratini, A. Private Communication
8) VanderHart, D. L., Wang, F. W., Eby, R. W., Fanconi, B. M. and DeVries, K. L.,
"Exploration of Advanced Characterization Techniques for Molecular
Composites", AFWAL-TR-85-4137, February 1986
110-18
1990 USAF-UES SUMMER FACULTY RESEARCH PROGRAM
GRADUATE STUDENT RESEARCH PROGRAM
Sponsored by the
AIR FORCE OFFICE OF SCIENTIFIC RESEARCH
Conducted by the
Universal Energy Systems, Inc.
FINAL REPORT
Potentials of Mushy-State Forming of Composite Materials
Prepared by:
Academic Rank:
Department and
University:
Research Location:
USAF Researcher:
Date:
Contract No:
Sherif D. EIWakil, Ph.D.
Professor
Mechanical Engineering
Southeastern Massachusetts University
WRDC/MLLM
Wright'Patterson AFB
Dayton OH 45433
Mr James C. Males
28 August 90
F49620-88-C-0053
by
Sherif D. Ei Wakil
ABSTRACT
Experimental work was carried out to investigate the problems involved in the
mushy-state forming of dispersion-strengthened composites, as weir as to assess the
possible potentials of such processes. Billets, all having the same Ai-Cu-Mn matrix but
different alumina contents, were obtained by hot compaction of canned powder
mixtures in a blind extrusion die. They were then homogenized for two hours before
being extruded at the required temperature. Three temperatures were chosen, to yield
different liquid fractions for the extrusion billets, namely 0.4, 0.2 and zero. Billets that
had 0.2 liquid fraction were successfully extruded, resulting in sound, defect-free
products. Also, metallographic examination of those mushy-state extruded bars
revealed an excellent degree of homogenity. In addition, the density and chemical
composition were found to be uniform along the length for those bars, indicating the
absence of any sensible segregation.
111-2
Achnowledgements
I wish to thank the Air Force Systems Command and the Air Force Office of Scientific
Research for sponsorship of this research. Acknowledgements must also go to the
Universal Energy Systems for their concern and great help to me iri all aspects of this
program.
My experience was extremely rewarding and enriching not only because of the
outstanding facilities of the Materials Lab at WPAFB, but also because of the
interaction with the highly qualified scientists in it. The encouragement of James
Malas and the fruitful discussions I had with him clearly added to every aspect of this
research. The help of Dr Venkat Seetharaman, Dr Young Kim and Carl Lombard was
invaluable in overcoming many technical difficulties. Thanks are also due to the staff
of the foundry lab, metallograpy lab, SEM lab. Chemical analysis Department, and
Mark Dodds. Last, but net least, Lwould like to thank Jim Morgan and Bill O'Hara for
providing a truly enjoyable working atmosphere.
I.
INTRODUCTION:
Mushy-state forming is gaining increasing attention as a potential process for
manufacturing complex shaped components using intermetallic and ceramic
composite materials.
The Material Processing Group of the Metals and Ceramics Division of the Materials
Laboratory at Wright-Patterson Air Force Base is particularly interested in developing
this new method. The goal is the production of light-weight composites with superior
mechanical properties that meet the demands of the aerospace industry in the
nineties.
My research interests have been in the area of mushy-state alloys. My work on
modeling of the plastic behavior of mushy-state alloys contributed to my assignment to
the Materials Processing Group.
111-4
II. OBJECTIVES OF THE RESEARCH EFFORT:
Currently, dispertion-strengthened composites are manufactured by casing or powder
metallurgy. Although these processes have some advantages, they still have some
limitations and shortcomings, for Instance segregation, nonhbmoginity, and
fragmentation of the hard particles. There is, therefore, a need to develop a new
manufacturing method to eliminate the above-mentioned shortcomings.
My assignment as a participant in the 1990 Summer Faculty Research Program
(SFRP) was to assess the feasibility of mushy-state forming processes for producing
Complex shaped components using intermetallic and ceramic composite materials.
Since time was a real constraint, it was decided to Investigate primarily the processing
parameters and to assess the problems associated with that process. An investigation
into the effect of the process parameters on the mechanical properties and their
correlation with the microstructure obtained, will be continued at my laboratory with
funding from the Mini Grant Prpgrartt.
111-5
III. EXPERIMENTAL WORK:
Experimental work was carried but to investigate the problems involved in the mushy'
state extmsion of dispersion*strengthened composites. An attempt to employ the
conventional powder metallurgy process to obtain suitable billets was not successful.
The compressibility and compactibility of Al-Cu powder mixtures containing different
percentages of alumina were extremely poor. Also, the compaction dies suffered from
excessive wear. It was, therefore, decided to obtain the billets by hot compaction of
canned powder mixtures in a blind die. The canned billets were then homogenized for
two hours at the required temperature before extrusion.
in the first extrusion experiment, the billet had a temperature .of lUO^F and a liquid
fraction of about 0.4. The resulting extruded bar was bulged and torn. This mode of
failure was attributed to the expansion of the liquid phase. It was, therefore, decided to
carry out extrusion at 1040°F so that the metal matrix of each billet would contain only
20% liquid phase. The speed of the extrusion ram was kept at the minimum. Sound
and defect-free products wer^; obtained. Next, the density and chemical composition
were determined at various points along the length of each extruded bar. They were
found to be uniform, indicating the absence of any sensible segregation.
Finally, in order to have a bases for comparison, some billets were extruded at 900®F
i.e. there was no liquid phase in any of those billets. Although the extrusions were
seemingly good, a decisive factor that determines the appropriate extrusion
temperature, is certainly the final mechanical properties of the extruded bars.
IV. RECOMMENDATIONS:
a. The feasibility of forming dispersion-strengthened composites have been basically
proven. Nevertheless, the effect of the process parameters on the mechanical
properties as well as the correlation between4hese properties and the micrpstructures
must still be irivestigated.
b. The results of the above-mentioned investigation can be used to optimize the
process of extruding composites in their mushy state.
c. Mathematical modeling of the behavior of mushy-state alloys is required, if those
new processes are to be developed to their full potentials.
d. A yield function and a flow rule must be developed first. The constitutive equation
can then be incorporated in a FE code, thus coupling the temperature distribution and
the plastic flow in order to rationally design tool profiles and anticipate any defects that
may occur in the final products.
e. An effort should be made to include the microstructure charateristics in the
constitutive equation.
f. it is also recommended to incorporate an analysis of possible chemical reactions in
the proposed study.
111-7
El Wakil, S.D,, A Model Study of Metal Forming in the Mushy State. Advanced
Technology of Plasticity. 1 984, Vol 1 , pp. 45-49.
Kiuchi, M., et.al., Annuals of the CIRP. 1987, Vol 36-1 , pp. 173.
Kiuchi, M. and S. Sugiyama, Application of Mushy-State Extrusion, Proc. of Conf. on
Extrusion. 1990, pp. 1-25.
1990 USAF-UES SUMMER FACULTY RESEARCH PROGRAM/
GRADUATE STUDEOT RESEARCH PROGRAM
Sponsored by the
AIR FORCE OFFICE OF SCIENTIFIC RESEARCH
Conducted by the
Universal Energy Systems, Inc.
FmAL.,fi^£QRJ
STRUCTURAL ANALYSIS OF POLYMER PRECURSORS WITH
POTENTIAL NONLINEAR OPTICAL PROPERTIES
Prepared by:
Academic Rank:
Department and
University:
Research Location:
USAF Researcher:
David A. Grossie, Ph.D.
Assistant Professor
Department of Chemistry
Wright State University
AFWRDC/MLBP
Wright-Patterson AFB
Dayton, OH 45433
Douglas S. Dudis, Ph.D.
Date:
28 Sep 90
Contract No:
F49620-88-C-0053
STRUCTURAL ANALYSIS OF POLYMER PRECURSORS WITH
POTENTIAL NONLINEAR OPTICAL PROPERTIES
by
David A. Grossie, Ph.D.
ABSTRACT
Single-crystal x-ray diffraction data was collected on two compounds having potential
nonlinear optical (NLO) properties, C23H29NOS and C3tH5oN202S3. Both compounds
crystalliri in triclinic crystal lattices, the first having cell constants of asl0.340(2),
b«l 1.632(1), c»8.894(3) A, a-97,18(2), ^*103.18(2), and y»88.05(1)«. The second
compound has cell constants of a«10.292(2), b*20.231(8), c«9.270(2) A, a»102.48(l),
pa98.67(2), and ^=88.66(1)'*. The space group observed in each compound is PI The
structure of compound 1 was solved and refined, yielding a R-factor of 0.061.
C23H29NOS is planar with little distortion in the internal bond distances and angles.
The second compound, C}gHjoN202S3, has not been completely solved, in spite of
application of the most recent and capable direct methods programs.
112-2
ACKNOWLEDGEMENTS
I would like acknowledge the Air Force Sy^ms Command and the Air Force Office of
ScienoHc Research for sponsorship of this research. Additionally, acknowledgement of
Universal Energy Systems must be given for their assistance in the administrative aspects
of the program.
I wish to thank the personnel of the Polymer Bnnch of the Materials Laboratory and
University of Dayton Research Institute for their assistance during the execution of this
research project.
112-3
1. INTRODUCTION:
The Polymer Branch of the Materials Laboratory at the Wright Aeronautical Laboratory,
Wdght-Patterson Air Force Base is interested in the synthesis and characterization of
polymeric materials. Basic research is also conducted in the structure of polymeric
materials and the correlation of the structure and physical properties, the emphasis of
this area is to predict the properties of a polymer prior to its synthesis. In this way, the
synthesis problem can haVe greater direction and produce new and better materials with
more efficiency. One of the techniques used in determining the structure of polymers is
to examine by angle-crystal x-ray diffraction methods compounds that may be used to
form the backbone, pendents, or cross-links of the polymer. By knowing the structure of
a small, repeating portion of the polymer, the polymer itself may be mathematically
modeled, yielding the physical properties.
My research interests are in the structural analysis of small organic and inorganic
molecules by single-crystal x-ray diffraction. My experience in the structural
determinations of organic molecules with similar features to those of interest and my
considerable familiarity with the available diffraction hardware and software contributed
to my assignment to the Morphology Section of the Polymer Branch.
112-4
n. OBJECTIVES OF THE RESEAROI EFFORT;
A study of nxxlel compounds of polymeric materials that have potential nonlinear optical
properties will be conducted. This study will involve the striictural analysis of several
compounds by single-crystal x-ray diffraction techniques, with the intent to amass data
which may be used to correlate the observed structure and the magnitude of the nonlinear
optical response. The primary structural information that is needed by the currently
accepted theories is the centricity of the crystal lattice in which the compound of interest
crystallizes and the extent of tc-orbital conjugation. To this end, this study will examine
the effects of steric bulk on the centricity observed in the crystal structure of an otherwise
plariu* molecule.
in.
a. Crystalline samples of a series of compounds were examined using an optical
microscope to determine the size and quality of the individual crystals. Two of the
compounds examined showed promise of containing suitable crystals for diffraction
analysis, whereas the remainder were of insufricient size to be analyzed. Single crys^s
of C23H29NOS ^d C3,H5oN202S3 were prepared for analysis by attaching them to a thin
glass fiber and pli^ing them at the center of an Enraf-Nonius CAD-4 automated
diffractometer.
112-5
Prelimin^ x-ray analysis of each of the select ciystals was made. These results are
summarized in Table 1, along with the parameters of the subsequent data coUecdon.
/
b. Data collected on each of the crystalline samples was examined for the presence of
space-group-deteimining systematic absences using the program LOOK (Qiapius, 1984).
For each sample, an appropriate space group was determined-PI for both compounds.
c. Using the program GENTAN (Hall, 29^), one of the many crystallographic routines
in the software package XTAL (Hall and Stewart, 1989; Grossie, 1990), the structure of
C23H29NOS was determined. The initial structure was refined using the full-matrix
least-squares refinement routine SFLSX (Hall, Spadaccini, Olthof-Hazekamp and Dreissig,
1989) contained in XTAL. Metrical details of the refined structure are tabulated in Tables
2-7. The structure of C3|H3oN202S3 was not completely solved, in spite of the application
of several of the most recent and capable direct methods programs available in the
laboratory. Work on tie solution of this compound is continuing.
d. Figure 1 shows an ORTEP (Johnson, 1971; Davenport, Hall and Dreissig, 1989)
drawing of the refined structure of C23H29NOS, and a summary of the interatomic
distances and angles is presentee! in Table 2. The central ring system of the molecule is
planar, with a miximum deviation from planarity of 0.044 A. The degree of pl^arity of
the central benzaie ring, 0.009 A and the benzothiazole ring system, 0.005 A, are much
better. This woild indicate that the two ring systems are slightly twisted with respect to
each other. This is also seen in the examination of the dihedral angles about the C2-C10
112-6
bond. The oxydecyl chain that is ortho to the benzothiarale ring is found to extend
slightly below the plane of the central benzene ring, with deviations increasing from 0.014
A for Oil to 1.370 A for C25.
e. Figure 2 shows an ORTEP drawing of die unit cell contents of C23H29NOS. The
packing of the molecules in the unit cell is in the herringbone pattern, a common feature
in the structure of planar compounds.
IV.
From the completed analysis of C23H29NOS and partial analysis of C3|HjoN202S3, certain
conclusions can be made regarding to the nonlinear optical properties of these compounds.
Since both crystallize in lattices that are centrosymmetric, the second-order term of the
polarizability equation must be zero. The response due to the third-order term for these
compounds is less certain. The effect of bulky substituents on the centricity of the crystal
lattice observed for a compound is still uncertain. The study of compounds with similar
structural features may provide the additional information to begin to understand the
influence of steric bulk in the arrangement of molecules into acentric lattices.
V. RECOMMENDATIONS:
There is a need to improve the synthesis of nonlinear optical materials and influence the
formation of acentric lattices. At the current time, many compounds show great potential
for high NLO responses, however the tendency is for them to cryst^lize in
centrosymmetric lattices. Synthetic modifications of these materials to favor the
foimation of noncentric lattices should be possible with the knowledge of the mimimum
perturbation to the compound necessaiy to promote asymmetry. The minimum
perturbation is needed so as to limit the effect of the change on the NLO response.
Currently, there are three factors that are assumed to produce the desired NLO response,
with one factor being quantitative and a second based on a relative scale. These two
factors are the centricity of the crystal lattice and the electron-donating and withdrawing
effects of the commonly used functional groups. The third factor, the extent of
conjugation within the molecule, is currently unquantifi^.
Molecular planarity is normally taken as the first clue that a non-fused, rc-bonded ring
system is conjugated. Since this information, like the centricity, is directly obtainable
from the structural analysis of a crystalline compound, the extent of conjugation can be
quantified by this process.
As a final step, the same compounds eximined by x-ray diffraction need to be tested for
a nonlinear optical response and the magnitude of that response. With the above pieces
of data obtained and analyzed, the synthesis of nonlinear optical materials can be by
rational design. This will allow the physical properties of the material to be optimized
without compromising the desired nonlinear optical properties.
112-8
REFERENCES
Chapius, G. (1984) "LCXDK. A FORTRAN Program for Generating Simulated Precession
Photographs from Diffractometer Data," University of Lausanne.
Davenport, G., Hall, and Dreissig, W. (1989) "ORTEP" XTAL 2.6 User’s Manual. Eds.
S. R. Hall and J. M. Stewart Universities of Western Australia and Maryland.
Grossie, D. A. (1990) "Desktop Crystallography: XTAL for the Personal Computer,"
Intemadonal Union of Crystallography Congress, Bordeaux, France.
HaU, S. R. (1989) "GENTAN" XTAL 2.6 User’s Manual. Eds. S. R. HaU and J. M.
Stewart. Universities of Western Australia and Maryland.
Hall, S. R., Spadaccini, N., Olthof-Hazekamp, R., and Dreissig, W. (1989) "SFLSX"
XTAL 2.6 User’s Manual. Eds. S. R. Hall and J. M. Stewart. Universities of Western
Australia and Maryland.
Hall, S. R. and Stewart, J. M. (1989) Eds. XTAL 2.6 User’s Manual. Universities of
Western Australia and Maryland.
Johnson, C.K. (1971) ORTEP n. Report ORNL-3794, revised. Oak Ridge National
Laboratoiy.
Table 1. Experiinental Details
Formula
Formula Weight
F(0(X))
Crystal Dimensions
Radiation
Wavelength
Temperature
Crystal Form
Space Group
Cell Constants
Volume
Z
Density
Absorption Coefficient
Scan Type
Scan Width
Maximum 2
Reflections Measured
Corrections
Observations
Parameters
R
wR
S
Maximum Shift/Error
Residual Density
Maximum
Minimum
Compound 1
QjHj^OS
367.55
396
0.2 X 0.3 X 0.4 mm
MoKa
0.71073
23‘’C
Triclinic
PI
a = 10.340(2) A
b= 11.632(1)
c * 8.894(3)
a « 97.18(2)*
P = 103.18(2)
7* 88.05(1)
1033.4
2
1.18 g/cm’
1.68 cm’*
co:20
1.00 + 0.344 taj 0
54.0*
4060
Numerical absorption
(range of correction =
Compound 2
0*^50^202^3
663.03
712
MoKa
0.71073
23*C
Tiiclinic
PI
a = 10.292(2) A
b = 20.231(8)
c = 9.270(2)
a = 102.48(1)*
P = 98.67(2)
7 * 88.66(1)
1863.0
2
1.18 g/cm^
2.22 cm’*
cd:20
0.85 + 0.344 tan 0
60.0
11492
L036 - 1.048)
Reflection avenging (Rint
* 0.69% for 481 duplicate
reflections)
1957
235
0.061
0.086
1.014
0.016
0.7 e/A'
-0.4 e/A^
Tabic. 2. Atom Codrdin^ and Isotropic Thennal Parameters for GjaHj^NQS.
x/a
y/b
z/c
U
S(l)
0.1231(1)
0.1319(1)
0.2004(2)
6.0756(6)
e(2)
0.1997(4)
0.0016(4)
0.1435(5)
0.069(2)
N(3)
0.3053(4)
0.0152(3)
6.0904(4)
0.073(2)
C(4)
0.3315(4)
0.1303(4)
0.0896(5)
0i072(2)
C(5)
0.4368(5)
0.1729(4)
0.0403(6)
0.086(3)
C(6)
0.4511(6)
0.2902(5)
0.0471(7)
0.096(3)
C(7)
0.3619(6)
0.3664(5)
0.1037(8)
0.100(3)
C(8)
0.2585(5)
0.3272(4)
0.1536(7)
0.092(3)
C(9)
0.2424(4)
0.2079(4)
0.1470(5)
0.074(2)
C(10)
0.1517(4)
-0.1130(4)
0.1531(5)
0.067(2)
e(ii)
0.0400(4)
-(). 1318(4)
a2122(5)
0.072(2)
0(11)
-0.0225(3)
-6.0361(3)
0.2638(4)
0.082(2)
C(12)
0.0005(5)
-0.2436(4)
0.2173(6)
0.086(3)
C(13)
0.0679(6)
-0.3363(5)
0.1607(7)
0.095(3)
G(14)
0.1765(6)
-0.3196(5)
0.1030(8)
0.100(3)
C(15)
0:2169(5)
-0.2090(4)
0.0968(6)
0.085(3)
C(16)
-0.1420(4)
^0.0473(4)
0.3176(6)
0.074(2)
C(17)
-0.1938(5)
0.0721(4)
0.3524(6)
0.077(2)
C(18)
-0,3186(4)
0.0726(4)
0.4152(6)
0.077(2)
C(i9)
-0.365^5)
0.1935(4)
0.4605(6)
0.079(2)
C(20)
-0.4899(5)
0.2002(4)
0.5240(6)
0.680(3)
C(21)
-0.5337(5)
0.3232(5)
0.5624(7)
0.093(3)
C(22)
-0.6630(5)
0.3359(6)
0.6121(8)
0.105(3)
C(23)
-0.7110(7)
0.4674(6)
0.6287(11)
0.137(5)
C(24)
-0.8276(8)
0.4906(7)
0.6734(11)
0.136(5)
C(25)
-0.8727(7)
0.6129(6)
0.6758(10)
0.124(4)
112-11
Tables.
H(5)
H(6)
H(7)
H(8)
H(12)
H(13)
H(14)
H(15)
H(16A)
H(16B)
H(17A)
H(17B)
H(18A)
H(18B)
H(19A)
H(19B)
H(20A)
H(20B)
H(21A)
H(21B)
H(22A)
H(22B)
H(23A)
H(23B)
H(24A)
H(24B)
H(25A)
H(25B)
H(25Q
Hydrogen Coordinate and Isotropic Thermail Parameters for GjjHj^OS;
x/a y/b
6.5053(5) 0.1182(4)
0.5248(6) 0.3217(5)
0.3765(6) 0.4524(5)
0.1947(5) 0.3838(4)
-0.0799(5) -0.2377(4)
0.0416(6) -0.4178(5)
0:2240(6) -0.3885(5)
0.2961(5) -0.1988(4)
-6.2118(4) -0.0967(4)
-6.1206(4) -0.0877(4)
-0.1193(5) 41222(4)
-0.2201(5) 0.1088(4)
-0.3948(4) 0.0265(4)
-6.2937(4) 0.0312(4)
-0.2897(5) 0.2382(4)
-0.3871(5) 0.2353(4)
-0.5659(5) 0.1540(4)
-0.4684(5) 0.1615(4)
-0.4593(5) 0.3663(5)
-0.5460(5) 0.3630(5)
-0.7375(5) 0.2894(6)
-0.6501(5) 0.3022(6)
-0.6395(7) 0.5092(6)
-0.7093(7) 0.5018(6)
-0.8991(8) 0.4439(7)
-0.8268(8) 0.4626(7)
-0.8012(7) 0.6596(6)
-0.8735(7) 0.6409(6)
-0.9656(7) 0.6350(6)
z/c
U
0.0059(6)
0.075
0.0067(7)
0.075
0.i083(8)
0.075
0.1942(7)
0:075
0.2588(6)
0.075
0.1683(7)
0.075
0.0590(8)
0.075
0.0511(6)
0.075
0.2317(6)
0.075
0.4197(6)
0.075
0.4307(6)
0.075
0.i475(6)
0.075
0.3339(6)
0.075
0.5164(6)
0.075
0.5439(6)
0.075
0.3595(6)
0.075
0.4427(6)
0.075
0.6275(6)
0.075
0.6497(7)
0.075
0.4603(7)
0.075
0.5289(8)
0.075
0.7187(8)
0.075
0.7208(11)
0.075
0.5255(1 i)
0.075
0.5863(11)
0.075
0.7813(11)
0.075
0.7628(10)
0.075
0.5678(10)
0.075
0.7005(10)
0.075
112-12
Table 4. Anisotropic Thenxial Displaccinenr Pax^eters for (^H^OS.
un
U22
U33
U12
U13
U23
S(1)
0.0718(7)
0.0744(7)
0.0848(7)
0.6099(5)
■0.0272(5)
0.0109(5)
C(2)
0.665(2)
0i076(3)
0.063(2)
0.008(2)
0.012(2)
0.009(2)
N(3)
0.072(2)
0.074(2)
0.076(2)
-0.001(2)
0.025(2)
0.004(2)
C(4)
0.073(3)
0.072(3)
0.076(3)
0.001(2)
0.016(2)
0.009(2)
C(5)
0.076(3)
0.084(3)
0.099(4)
-0.602(2)
6.027(3)
0.008(3)
C(6)
0.082(3)
0.091(3)
0.123(4)
-0.004(3)
0.034(3)
0.027(3)
C(7)
0.099(4)
0.680(3)
0.125(5)
-0.007(3)
0.025(3)
0.025(3)
C(8)
0.086(3)
0.076(3)
0.114(4)
6.009(2)
0.023(3)
0.015(3)
C(9)
0.067(3)
0.079(3)
0.077(3)
.0.004(2)
0;015(2)
0.014(2)
C(10)
0.066(2)
0.071(2)
0.065(2)
-0.001(2)
0.015(2)
0.006(2)
C(ll)
0.070(2)
0.072(3)
0.071(3)
0.001(2)
0.018(2)
6.004(2)
0(11)
0.082(2)
0.076(2)
0.100(2)
0.001(1)
0.044(2)
0.006(2)
C(12)
0.081(3)
0;083(3)
0.099(3)
-0.004(2)
0.032(3)
0.009(2)
C(13)
0.102(4)
0.075(3)
0.113(4)
-0.004(3)
0.037(3)
0.016(3)
C(14)
0.101(4)
0.074(3)
0.130(5)
0.005(3)
0.049(3)
-0.004(3)
C(15)
0.087(3)
0.080(3)
0.094(3)
0.005(2)
0.036(3)
0.004(2)
C(16)
0.060(2)
0.081(3)
0.082(3)
-0.004(2)
0.021(2)
0.008(2)
C(17)
0.070(3)
0.089(3)
6.078(3)
0.002(2)
0.027(2)
0.012(2)
C(18)
0.064(2)
0.081(3)
0.087(3)
-0.003(2)
0.019(2)
0.010(2)
C(19)
0.067(3)
0.091(3)
0.081(3)
0.063(2)
0.022(2)
0.013(2)
C(20)
0.067(3)
0.087(3)
0.087(3)
0.001(2)
0.022(2)
0.007(2)
C(21)
0.076(3)
0.098(4)
0.107(4)
0.002(3)
0.027(3)
0.003(3)
C(22)
0.077(3)
0.117(4)
0.117(4)
0.005(3)
0.026(3)
-0.015(3)
C(23)
0.110(5)
0.131(5)
0.170(7)
-0.032(4)
0.061(5)
-0.049(5)
C(24)
0.115(5)
0.128(6)
0.169(7)
-0.011(4)
0.042(5)
0.008(5)
C(25)
0.113(5)
0.098(4)
0.161(6)
0.008(4).
0.038(4)
0.013(4)
112-13
Tabic 5. Bond Distances (A) and Angles (*) for C23H29NOS.
Atoms
Distance
Atoms
Angles
S(l)-C(2)
1.758(4)
G(2)-S(l)-G(9)
89.4(2)
S(l)-C(9)
1.724(5)
S(l)-G(2)-N(3)
114.3(3)
CC2)-N(3)
1.307(6)
S(1)-G(2).G(10)
124:0(4)
C(2)-C(10)
1.457(6)
N(3).G(2)-C(10)
121.7(4)
N(3)-C(4)
1.377(6)
G(2)-N(3).G(4)
111.9(4)
C(4)-C(5)
1.390(8)
N(3)-G(4)-G(5)
125.7(4)
C(4)-C(9)
1,403(7)
N(3)-G(4).G(9)
114.7(4)
C(5)-C(6)
1.370(8)
G(5)-G(4).G(9)
119.6(4)
C(6)-C(7)
1.391(9)
G(4).G(5)-G(6)
119.4(5)
C(7)-C(8)
1.360(9)
G(5)-G(6)-G(7)
120.6(6)
G(8)-C(9)
1.395(7)
G(6)-G(7)-G(8)
121.3(5)
C(10)-C(il)
1.408(7)
G(7)-G(8).G(9)
118.8(5)
C(10)-G(15)
1.384(7)
S(l)-G(9)-G(4)
109.7(4)
G(ll)-0(11)
1.352(5)
S(l)-G(9)-G(8)
130.0(4)
G(ll)-G(12)
1.385(7)
G(4)-G(9)-G(8)
120.3(5)
0(11)-C(16)
1,437(6)
G(2)-G(10)-G(ll)
123.7(4)
G(12)-G(13)
1.369(8)
G(2).G(ld)-G(15)
118.4(4)
G(13)-G(14)
1.365(10)
G(ll)-G(10)-G(15)
117.9(4)
G(14)-G(15)
1.378(8)
G(10)-G(ll)-O(ll)
116.3(4)
G(16).G(17)
1.493(7)
G(10).G(11).G(12)
120.3(4)
G(17)-G(18)
1.519(7)
0(11).G(11).G(12)
123.4(5)
G(18)-G(19)
1.510(7)
G(n)-0(11)-G(16)
119.9(4)
G(19)-G(20)
1.513(7)
G(11).G(12)-G(13)
120.0(5)
G(20)-G(21)
1.508(7)
G(12)-G(13)-G(14)
120.5(5)
G(21)-G(22)
1.499(9)
G(13)-G(14)-G(15)
120.2(5)
G(22)-G(23)
1.591(10)
G(10)-G(15)-C(14)
121.1(5)
G(23)-G(24)
1.362(12)
0(11)-G(16).G(17)
107.4(4)
G(24)-G(25)
1.481(10)
G(16)-G(17)-G(18)
112.7(4)
G(17).G(18)-G(19)
112.6(4)
G(18)-G(19)-G(20)
115.3(4)
G(19)-G(20)-G(21)
112.6(4)
G(20)-G(21)-G(22)
115.3(5)
G(21)-G(22).G(23)
111.9(6)
G(22)-G(23)-G(24)
117.7(7)
G(23)-G(24)-G(25)
115.5(8)
112-14
Tabic 6. Dihedral Angles O for CaH^NOS.
Atoms
Angle
Atoms
Angle
C(9)-S(l)-C(2)-N(3)
-0.3(3)
C(2)-C(10)-C(ll)-C(12)
-179.9(4)
C(9)-S(l)-C(2)-C(10)
179.6(4)
C(15)-C(10)-C(ll)-O(ll)
178.9(4)
C(2)-S(l)-C(9)-G(4)
-0.2(3)
C(15)-C(10)-C(ll)-C(12)
-2.0(Q
C(2)-S(l)-C(9)-C(8)
180.0(5)
C(2).,C(10)-C(15>C(14)
-179.8(4)
S(1VC(2).N(3)-C(4)
0.7(4)
C(U)-C(10)-C(15)-C(14)
2.1(7)
C(10)-C(2)-N(3).C(4)
-179.2(4)
C(10)-C(ll)-O(ll)-G(16)
-176.4(4)
S(1)<C(2)-C(10)-C(11)
1.4(6)
G(n)-G(ll)-(D(11)-G(16)
4.6(6)
S(l)-C(2)-C(10)-C(15)
-176.6(3)
G(10).G(11)-G(12)-G(13)
1.9(7)
N(3)-C(2)-C(10)-C(ll)
-178.8(4)
0(11)-G(11)-G(12)-G(13)
-179.1(4)
N(3)-C(2)-C(10)-C(15)
3.3(6)
G(ll)-0(11)-G(16)-G(17)
174.8(4)
C(2).N(3)-C(4)-C(5)
-179.8(4)
G(ll)-C(12)-G(13)-C(14)
-1.9(8)
C(2)-N(3)-C(4)-C(9)
-0.9(5)
G(12).C(13)-G(14)-G(15)
2.0(9)
N(3)-C(4).C(5)-C(6)
179.6(5)
G(13)-C(14)-G(15).G(10)
-22(8)
C(9).C(4).C(5)-C(6)
0.7(7)
0(11)-G(16).G(17)-G(18)
178.3(3)
N(3)-C(4)-C(9)-S(l)
0.6(5)
G(16)-G(17)-G(18)-G(19)
-176.1(4)
N(3)-C(4)-C(9)-C(8)
-179.5(4)
G(17)-G(18).G(19)-G(20)
-180.0(4)
C(5)-C(4).C(9)-S(1)
179.7(4)
G(18)-G(19)-G(20)-G(21)
178.3(4)
C(5).C(4).C(9)-C(8)
-0.5(7)
G(19)-G(20)-G(21)-G(22)
•174.7(5)
C(4)-C(5).C(6).C(7)
-0.4(8)
G(20)-G(21)-G(22)-G(23)
172J(5)
C(5).C(6)-C(7)-C(8)
0.0(9)
G(21)-G(22)-G(23)-G(24)
180.0(7)
C(6)-C(7)-C(8)-C(9)
0.2(9)
G(22)-G(23)-G(24)-G(25)
175.6(7)
C(7)-C(8)-C(9).S(1)
179.9(4)
C(7)-C(8)-C(9)-C(4)
0.0(7)
C(2)-C(10)-C(ll)-O(ll)
1.0(6)
112-15
Table 7. Least Squares Planes for C23H29N6S.
Plane 1 0.5960 X + -0.0066 Y + 0.8030 Z = 1.8501
Atom Deviation Atom Deviation Atom Deviation Atom Deviation
CIO 0.006(6) Cll r0.007(6) C12 0.007(7) C13 -0.005(8)
C14 0.079(8) C15 0.009(7)
— Other Atoms —
Atom Deviation Atom Deviation Atom Deviation Atom Deviation
on -0.014(7) C16 -0.106(10) C17 -0.219(12) C18 -0.284(15)
C19 -0.3(j4(18) C20 -0.369(21) C21 -0.427(24) C22 -0.619(27)
C23 -0.861(30) G24 -1.050(33) C25 -1.370(35)
Plane 2 0.5730 X + 0.0248 Y + 0.8192 Z » 1.8128
Atom Deviation Atom Deviation Atom Deviation Atom Deviation
SI 0.000(2) C2 -0.002(5) N3 0.004(5) C4 -0.005(5)
C5 0.001(6) C6 -0.001(7) a 0;^i00(8) C8 0.003(7)
C9 0.001(5)
Plane 3 0.5730 X + 0.0248 Y + 0.8192 Z = 1.8128 '
Atom Deviation Atom Deviation Atom Deviation Atom Deviation
SI -0.003(2) C2 0.014(5) N3 0.025(4) C4 0.004(5)
C5 0.012(6) C6 -0.002(7) C7 -0.015(7) C8 -0.015(7)
C9 -0.004(5) CIO 0.014(5) Cll 0.028(5) C12 0.025(6)
C13 -0.029(7) C14 -0.043(7) C15 -0.044(6)
Angles Between Planes
Plane Plane Angle Plane Plane Angle
1 2 2.4(2) 1 3 1.8(1)
2 3 0.6(1)
112-16
112-17
1990 USAF-UES SUMMER FACULTY RESEARCH PROGRAM/GRADUATE
STUDENT RESEARCH PROGRAM
Sponsored by the
AIR FORCE OFFICE OF SCIENTIFIC RESEARCH
Conducted by the Universal Energy Systems, Inc.
FINAL REPORT
Eddy CurrenLTesting in Nonde-structivs. Evaluatiofl
Prepared by:
Academic Rank:
Department and
University:
Research Location:
Thomas J. Haas, B.S.E.E., B.S. Physics
P.K. Kadaba, PhD.
Graduate Student (M.S.E.E.)
Professor, University of Kentucky
Electrical Engineering
University of Kentucky
WRDC/MLLP
WPAFB
USAF Researcher:
Date:
Contract No:
Dayton, OH 45433
P.K. Bhagat
31 Aug 90
F49620-88-C-0053
Eddy Current Testing in Nondestructive Evaluation
by
TJ. Haas and P.K. Kadaba
ABSTRACT
Advantages and limitations of the eddy current technique for the
purpose of nondestructive testing have been evaluated. A cursary study of some
of the analytical models and actual test systems that have been developed by
researchers over the years has been made.
Using the commerciaiiy available eddy current testers - the Nortec
NDT-16 and the Hocking AV100SE • tests were made on samples of rubidium
and stainless steel with standard machined notches of depths 0.2mm, 0.4mm,
and 1 .0mm. Also tested was an unknown sample with a bareiy visible crack.
The HP4192A Impedance Analyzer was adapted to eddy current testing by
incorporating a power amplifier and specially designed transmit and receive
probes. A minute hole in a sample of aluminum was easily detected with this
set-up. A pulse technique capable of detecting defects in nonmagnetic metals to
a depth of 0.5cm or better was developed. This technique seems to have
potential to detect second layer cracks.
113-2
ACKNOWLEDGEMENTS
We wish to thank the Air Force Systems Command as well as the Air Force
Office of Scientific Research for the opportunity to work with quality people in a
quality environment.
Specifically, we would like to make special mention of the efforts of Dr. P.K.
Bhagat, Dr. Thomas J. Moran, and“Mr. Mark Blodgett, all of whom guided us with
their research expertise. The talents of Mr. Mark Ruddell and Mr. Ed Klosterman
need to be recognized, as well as Mr. J.A. Fox and Ms. L.L. Mann. The efforts of
Mrs. Nancy Lammers were also appreciated.
Of course Universal Energy Systems needs mentioning for organizing and
administrating this opportunity.
113-3
I. INTRODUCTION:
Eddy current testing is used in the fiei'^ of nondestructive evaluation to perform
the inservice inspection of metal (conductor) products. The experienced
operator can, in general, determine the presence of surface flaws (subsurface
flaws in nonmagnetic conductors), layer thickness, and perform metal sorting
(conductivity).
The method is used to perform these types of inspections because they can be
carried out quickly with no contamination of the surface. Despite the simplicity of
the technique, it should be noted that the inversion problem, that is, obtaining
detailed characterization of surface and subsurface defects of a test object from
the probe output signal is a complex problem. Many factors influence the output
signal: frequency of excitation of the probe coil, electrical conductivity and
magnetic permeability of the material, geometry of the test object and search
coil, as well as discontinuities and inclusions in the material.
II. OBJECTIVES OF THE RESEARCH EFFORT:
In the brief period of the program, we have restricted our effort to the evaluation
of the following procedures. These are (i) use of Nortec NDT-16 and Hocking
AV100SE eddy current testers, (ii) use of the HP4192A Impedance Analyzer, (iii)
a Hall Effect pulsed eddy current tester - Nortec-30 Eddyscan, and (iv) a
modification of a pulse techniqje reported in the literature. The above decision
was motivated by the availability of the in house commercial equipment in the
first two cases. Evaluation of the Hall Effect device was undertaken as it has
potential to detect subsurface cracks to about 40 mils in metallic conductors.
113-4
The pulse technique researched by us seems promising and should be
applicable to investigation of in-depth defects and also for investigation of
magnetic materials.
The report is organized as follows: after a brief;|iccount of basic principles and
analytical models reported in the literature, the four experimental techniques
proposed have been described in some detail. This is followed by the lab
results obtained by us and the report concludes with a brief discussion of the
results.
Hi. THEORETICAL CONSIDERATIONS
a. ) An extensive look at the literature provided insight into the basic principles of
the eddy current technique. Textbooks exist on the technique specifically (as
well as on nondestructive evaluation (NOE) in general), oftentimes dedicating
chapters to the subject at hand; A number of articles in the various journals of
the NDE field were sought out as well. These are listed in the bibliography at the
end of the report.
b. ) The eddy current test employs some of the most basic principles of electricity
and magnetism that have been known to us for some 1 50 plus years. When an
object to be tested is placed in a varying magnetic field, varying currents are
induced in the test object. The original magnetic field (or primary field) is
produced by applying a varying current to a coil, the test probe of the eddy
current instrument. Those induced currents in turn produce a field of their own
(secondary field) which opposes the primary field. The “size” of this opposition
depends on the material characteristics of the tested object, conductivity,
113-5
frequency of the applied, current, arid physical dimensions of the test object and
probe coil.
the reduction in magnetic flux due to the opposing fields causes a change Ip
coil impedance, which is monitored. When the eddy current flow is disturbed,.as,
for example, by a defect, a resultant change in the magnetic field occurs, which
in turn affects the coil. impedance as described above.
From another viewpoint, when a probe is in air, the coil has a fixed inductance
and resistance. When the probe approaches a conducting surface, the apparent
resistance increases (due to losses in the conductor) and the inductance
decreases due to the oppd )ing fields. Inductive reactance and resistance
combine to yield impedance, a vector quantity (or phasor). Reactance is usually
assigned as the ordinate (vertical axis) and resistance as the abscissa
Inductive
Reactance
Resistance
Fig. 1. Impedance plane diagram.
Fig. 2. AC bridge circuit.
(horizontal axis) in an impedance plane diagram. The impedance plane
diagram is normally used as the display in presenting the eddy current test
information.
113-6
In a passive probe (coil), the probe impedance, whose dynamic characteristics
were described earlier, changes due to the presence of a test, object (and any
alterations therein). These changes are detected in a bridge circuit, translated
into a voltage, and then displayed on a CRT screen. This display is the
representation of the (complex) impedance plane.
Please note that Figure 1 is consistent with the coil impedance having a larger
resistive component and smaller reactive component as the probe approaches
the test object.
Impedance can be written in its complex form as Z = R + jX, where R represents
the resistive component and X is the inductive reactance component. A
dependence of impedance on frequency can be uncovered if we expand X as X
-<iX. = 27ifL, where L is coil inductance and f is the frequency of operation of the
excitation current.
Frequency of operation is one of the most important parameters in eddy current
testing, and is one of the few that the inspector has control over. The more
choices the inspector has in picking a frequency for the test (i.e. wider
instrument bandwidth), the more powerful the instrument, in general.
In the development of analytical models for eddy current characterization of
defects (voids) the integral equation approach has been widely used. A major
drawback in using the integral equation approach as the heart of a defect
characterization scheme lies with the basic dipole assumption: how can such a
model be used as the basis of the inverse problem solution when one has
113-7
already assumed a priori that the defect has a spheroidal shape? But this is what is
dohCi Work should be continued in this area for arbitrarily shaped defects.
As ah example of the approach taken in forinulating a finite element analysis, consider
a two dimensional region with z directed cuirent density vector J, For this geometry,
the magnetic vector potential A is directed parallel to i aiid is described by:
d
,3A
v-^
dx
dx
dy
dy
\ ^ J
=-l+yo)oA
(1)
where sinusoidal excitation of angular frequency co is assumed.
The principles of variational calculus lend a solution to (1) by minimizing
9A
R
V2
dx
[Sy]
dx dy
(2)
where R is the bounded tegion of interest. F is called the energy functional.
Analytical models for layer^ media revolve around finding a solution for the magnetic
vector potential, A as well. For a circular coil electrically close to a plane material
medium (dielectric constant Ej, conductivity gj, and permeability |ii) at z=0 and coil
axil in the z direction only the (cylindrical coordinates p, (j>, z) component has any
significance [Ref. 7].
By solving the wave equation and implementing the boundaiy conditions we find, for
z>0
2
0 ^
(3)
I n H-l^O ^
where Ji(ep) refers to the first order Bessel function of the first kind, d is the
113-8
coil/material separation, a is the coil radius, and X,- refers to the wavelength in region i
(i = 1 or 0 here). The term is brackets is analagous to the reflection coefficient expres¬
sion used in describing electromagnetic wave interaction with materials.
If the medium above is replaced by a layered media of thickness, t, Eii Hj, Oj and
E2, |i2’ ^2 describing its material propenies, we obtain A ^ as in (3) upon replacing the
term is brackets with
^ roi + ri,e-^'
where
r - . p _ M’2^1
+ 42^1+1^1^2
Upon making this substitution one obtains the solution of the magnetic potential.
For defect characterization work, a model is required which allows for a variety of
defect shapes, test geometries and excitation condition. Both flnite-difference and flnite
element techniques appear to have this flexibility. Even with these methods, published
work in the literature, by and large, is restricted to characterization of defects which
have well defined symmetries in nonmagnetic materials. Additional work must be done
to extend the techniques to to three dimensions as well as defects of arbitrary shapes
and to take account of nonlinear material properties so that defects which include hys¬
teresis effects in ferromagnetic materials can be studied. As with any computer-based
modeling, finite element analysis techniques give results whose accuracy is very much a
function of the quality of the input data.
113-9
IV. ACTUAL TEST SYSTEMS
a. ) Experiments were performed using the Nortec NDT-1 6 and the Hocking
AVtOOSE eddy current testers, and the HP4192A Impedance Analyzer.
Samples of Rhobidium and stainless steel with machined notches of depths
0.2mm, 0.4mm, and 1 .0mm were used, as well as an unknown sample with a
barely visible crack. These were analyzed to varying degrees by the
aforementioned devices. There were some limitations to the test procedure that
we could not undertake due to lack of calibration standards for use in eddy
current testing in the MLLP section. See section V. RECOMMENDATIONS, in
this report.
b. ) There was some difficulty In learning to use the Nortec NDT-1 6 eddy current
tester, as there was no documentation available on the instrument. By speaking
with individuals with some experience in eddy current testing and by using other
testers, a knowledge of this particular device was obtained.
By placing the probe on the test object and nulling the bridge, a new reference
point (or origin) for the impedance plane diagram is made. Now, as the surface
of the test object is scanned, changes in its materiai properties are monitored,
and the signal from these changes is with respect to the reference point. This is
vital in that the lift off signal (the signal change between probe in air and probe
on the test object) dominates the defect signals. So the probe is placed on the
material, nulled, and then the lift off signal is directed horizontally (by
113-10
convention), through the use of a phase rotation control,, ■I;hia:pfp.vi,des some
discrimination against the lift off signal. Now the smaller signals are
distinguished from. the lift off signal by both amplitude (smaller) and phase. Skin
depth is inversely proportional to the square root.pf frequency and conductivity.
The higher the frequency, the snialleTthe distance the currents penetrate- into
the material (for a constant conductivity). Currents at greater depths below the
material surface flow at greater phase angles that lag behind the currents
nearer the surface. Therefore, the signals generated from the lower frequency,
deeper penetrating currents lag behind signals generated by the higher
frequency currents. This can be observed in Fig. 4 and 5. A stainless steel
sample was scanned at constant gain for frequency, f=500kHz (Fig. 4) and
fsIMHz (Fig. 5). Note that the liftoff signal is in the horizontal direction and that
despite changes in amplitude for the three crack signals (amplitude proportional
to crack depth), an obvious phase difference is observed.
There are some trends present in the figures which remain consistent for any
sample (a rubidium sample machined exactly as the stainless steel sample was
also tested). They are: signal amplitude increases with frequency and crack
depth, and phase difference from liftoff signal increases with frequency and
crack depth.
Important limitations exist in using the Nortec NDT-16. The most important of
these is the frequency bandwidth, and the fact that only 3 discrete frequencies
are available. It would be greatly desirable to have continuous frequency
settings, as well as the capability to use frequencies well below the lower limit
on this instrument of 500kHz. This is a necessity in having the ability to detect
subsurface defects.
113-11
FIG. 4. Eddy current test signal for stainless steel
calibration standard taken from the Nortec NDT-1 6
at 500kHz.
FIG. 4. Eddy current test signal for stainless steel
calibration standard taken from the Nortec NDT-1 6
at 1MHz.
For the detection of subsurface cracks, low frequency techniques heed to be
implemented,. In attempting to inspect second and third layer materials, greater
success will be achieved if the outer layers are, in the case of a change in
material with layer, lower in conductivity. In lowering frequency, a necessary
increase in probe diameter is required. This results in an increase in the size of
detectable cracks.
Due to a lack of proper calibration samples, no scientific data was obtained for
subsurface cracks for this report. However, using the Hocking AV100SE
instrument, and simulating a crack by placing two thick (electrically) copper
sheets adjacent to one another and covering them with a sheet of metal
composite, the area between the copper sheets was an observable signal.
The HP Model 4192A LF Impedance Analyzer is a fully automatic, high
performance test instrument designed to measure a range of impedance
parameters as well as gain, phase, and group delay. Other features of the
instrument pertinent to the project are: -1 .) frequency of the oscillator output can
be varied from 5Hz to 13MHz, 2.) oscillator level is variable from 5m V to 1.1V
(rms) with ImV resolution, 3.) measurement range of gain/loss of the test
channel with reference to the reference channel is -lOOdB to +100dB with
0.001 dB maximum resolution and 0.02dB to 0.09dB basic accuracy, and 4.)
measurement range of phase is -180 to +180 degrees with 0.01 degrees
resolution and 0.1 degrees to 0.2 degrees basic accuracy. The 4192A provides
HP-IB interface capability and this feature makes it possible to integrate the
4192A into an automated test system which reduces time and cost of testing.
113-13
Fig. 3. Eddy current testing using the HP 4192A Impedance Analyzer.
In our tests the HP4192A was modified using an external power amplifier and
specially designed transmit and receive probes, for the eddy current
V
measurements. A sample of aluminum foil, 25.4 microns thick with a hole of
1 .07mm diameter was used in this setup (Fig. 3) with an excitation frequency of
1MHz. A 9.8dB change in the receiver probe signal magnitude was observed as
the sample was passed over the probes. It is to be noted that operating the coils
at their resonant frequency makes eddy current measurements difficult. Good
sensitivity was achieved in measuring the coil impedance by the impedance
analyzer. The resonant condition occurred well below 1MHz. This method
shows good potential for automation, and we plan to continue the study at the
University of Kentucky.
A method similar to that of Waidelich [Ref.7], with some modifications, was
investigated. This is a pulse technique and capable of detecting defects in
nonmagnetic metals to a depth of 0.5cm or better. A brief description of the
experimental setup used follows: the output of the HP8111A pulse generator
113-14
(prf: 1kHz) was amplified by a broadband power amplifier and fed to a
transmitting probe coil. The transmitting coil consisted of a 600 turn coil of
copper wire wound on a straight ferrite core of 3/16 inch diameter cross section.
Likewise the receiving probe coil was made of 2000 turns of wire wound on a
similar ferrite rod. The output of the receiver probe was amplified and displaced
on an oscilloscope. Both the probes were shielded axially to prevent leakage of
stray flux. Two of the important parameters are the height of the peak voltage of
the displayed pulse and the time delay from the beginning of the pulse to the
peak. For aluminum sheets these two quantities were measured for thicknesses
of 0.2 to 10 cms. It was found that the peak height varies as l/tss where t is the
thickness of the specimen and the time delay of the output pulse peak varies as
h-8. By moving the probe coils over the surface of the. sheet and observing the
behavior of the tail of the output pulse, the position and depth of the defects can
be ascertained. This technique has potential and it is propo.sed to continue
these measurements at the University of Kentucky. Preliminary tests on holes
drilled in one aluminum sheet and sandwiched between two other sheets seem
encouraging.
V. RECOMMENDATIONS
There is at present a lack of calibration standards for doing eddy current testing
at the MLLP branch. EDM notches or machined slits should be obtained for the
materials for which there is interest in inspecting. These calibration standards
should duplicate test material in geometry as well as in electrical and magnetic
properties. Due to the failure to characterize in detail the flaw signal in eddy
current analysis of materials, eddy current testing is not an absolute method.
113-15
The Nortec NDT-1 6 eddy current instrument currently in the MLLP branch has
serious limitations in probe excitation frequency as well as probe types and
sizes. There are instruments available that will go much lower in frequency. This
capability is a necessity, especially when materials classified as good
conductors are being evaluated for flaws. It becomes imperative if multi-layered
structures are to be analyzed. Probe (coil) size offers another choice to the
operator of test equipment, and a variety of coil diameters should be made
accessible.
It is our feeling that complete characterization of defects will combine
knowledge obtained from analytical and numerical models integrated with a
data base of signals obtained from calibration standards. Neural networks could
very well play the role of integrator. Calibration standards should include
signals from both machined defects as well as actual, typical defects of the
material to be inspected.
In conclusion, eddy current testing has proven to be reasonably successful in
the past. Further improvements in methodology and test equipment would
make the technique more versatile in flaw detection in metallic materials.
113-16
REPERB^C£S
RUDLIN, J.R., “A Beginnerfs Guide to-Eddy Current Testing." British
Journal of NOT. University College Underwater NDE Centre, June
1989, pp. 314-320.
Hagemair, D.J., “Eddy Current Impedance Plane Analysis," presented
to 1983 Air Transport Association Non-Destructive Testing Forum,
Atlanta, Georgia, September 1982, pp. 1-5.
Dodd, C.V., Whitaker, LM., and Deeds, W.E., “An Accurate Laboratory
Test System Using Commercial Equipment for Eddy Current Measure¬
ments," .MatfijialS-Eyallialion. November 1988, pp. 1569-1574.
Libby H.L, Introduction to Electromagnetic Nondestructive Test
Methods. Richland, Washington, Wiley-Interscience Publishers, 1971.
Cecco, V.S., and Van Drunen, G.,"Recogni2ing the Scope of Eddy Cur¬
rent Testing," in Research Techniques in Nondestructive Testing YqI. VIII
Sharpe, R.S., Ed., Academic Press Inc., London, 1985, pp. 270-301.
Auld, B.A., Muennemann, F.G., and Riazat, M., “Eddy Current Model¬
ling," in Research Techniques in Nondestructive Testing Vol. VII.
Sharpe, R.S., Ed., Academic Press Inc., London, 1984, pp. 37-76.
Waidelich, D.L, “Pulsed Eddy Current Testing of Steel Sheets," in
Eddy,.Curcgnt CharacterizatlQD-QiJi^atsiials.and., Structures. Birnbaum/
Free, Ed., ASTM Special Technical Publication 722, Philadelphia,
1981, pp. 367-373.
Palanisamy, R., and Lord, W., “Finite Element Analysis of Eddy Current
Phenomena," Materials Evaluation. October 1980, pp. 39-43.
Cheng, D., “The Reflected Impedance of a Circular Coil in the
Proximity of a Semi-Infinite Medium," PhD Dissertation, University
of Missouri, January 1964.
113-17
1990 USAF-UES SUMMER FACULTY RESEARCH PROGRAM,'
GRADUATE STUDENT RESEARCH PROGRAM
Sponsored by the
AIR FORCE CFHCE OF SCIENTIFIC RESEARCH
Conducted by the
Universal Energy Systems, Inc.
Prepared by: Joseph B. Lambert
Academic Rank: Professor
Department and Chemistry
University: Northwestern University
Research Location: WRDC/MLPJ (Electromagn. Mat. Div.)
Wright-Patterson AFB
Dayton, OH 45433-6539
USAF Researcher: Robert L. Crane
Date: June 15, 1990
Contract: F49620-88-C-0053
Preparation and Characterization of Polypeptide Thin Films
by
Joseph B. Lambert
ABSTRACT
Polypeptides based on the glutamic acid backbone have nonlinear
optical properties. In order to test practical applications of such
materials, thin films have been cast by use of spin coating. Films
were prepared for poly(benzyl-t-glutamate) (PBLG) and for polydL-
(p-tran s -azobenzeneVL-glUtamide’l (PALG). These films were
characterized by polarized microscopy, spectroscopic ellipsometry,
and Fourier transform infrared spectroscopy.
114-2
Acknowledgments
This work was sponsored by the Air Force Systems Command
and the Air Force Office of Scientific Research and was administered
by Universal Energy Systems, to whom the author is grateful. This
work was carried out in the Biotechnology Laboratory of the ivILPJ
branch of the Electromagnetic Materials Division of the^ Materials
Laboratory of Wright-Patterson Air Force Base. The author is
particularly indebted to Dr. Robert Crane, Dr. Thomas Cooper, Dr.
Zbigniew Tokarski, and Mr. Timothy Running for guidance and
assistance in this work.
114-3
I. INTRODUCTION
Organic nonlinear optical materials are of interest to the Air
Force for a wide variety of applications, including optical computing,
optical storage, wave guiding, and optical sensor/vision protection.
The major effort in recent years has been on highly conjugated
carbon chains such as polyacetylenes. The discovery that
polysilanes have NLO properties (Miller and Michl, 1989) revealed
the existence of an entirely new NLO material based on saturated
chains of silicon atoms. Electron delocalization in these materials is
based on a-^a* conjugation rather than the traditional n-¥n*
conjugation.
For the past ten years our research at Northwestern has centered
around the preparation and investigation of organic molecules
containing silicon, germanium, or tin that possess novel electronic
properties. We have compared the abilities of these three atoms to
delocalize charge through conjugation (Lambert et al., 1989)
and found that silicon and germanium have similar abilities but that
tin has a vastly enhanced ability. Consequently, we became
interested in the possibility that polystannes could have enhanced
NLO properties compared with those of polysilanes.
Polysilanes have many outstanding properties for optical uses,
including fast response times, low visible absorptivity and
scattering, good thermal stability, and good processibility.
Moreover, their are in the range of the better NLO conjugated
carbon chains. We expect polystannanes to possess all these positive
properties and in addition have greatly enhanced values of
because of the much hiper polari^bility of tin.
To learn about the preparation and processing of NLO polymers^ I
felt that experience in the Materials Laboratory Of Wright-PaUerson
Air Force Base would be invaluable* as that laboratory has a long
history of investigation of NLO materials. Consequently, I requested
this laboratory in order to gain exposure to materials methodology
and to work with NLO polymeric materials.
114-5
11. OBJECnVES OF THE RESEARCH EFFORT
The major objectives of this work were to obtain a thorough
grounding in the theory of nonlinear optics, to leaTn how to
manipulate long chain NLO polymers, and to understand what the
significant problems in NLO chemistry are today. In order to
achieve these objectives, a protocol was selected to prepare and
characterize a family of high polymers with NLO properties, the
derivatives of poly glutamic acid. Levine and Bethea (1976) have
measured the second-order and third-order nonlinear optical
properties of the a-helix forni of poly(benzyl-L-glutamate) (PBLG)
with molecular weight about 550 kilodalton. Suitable alteration of
the side chain might improve the NLO properties, so Dr. Thomas
Cooper of the MLPJ laboratory has been engaged in a synthetic
project to prepare such derivatives. Past work has concentrated on
studying the NLO properties of these materials in solution. It was
our specific objective to determine whether homogeneous thin films,
free of solvent, could be cast and characterized. Consequently, we
developed methods using the spin coating technique to prepare such
thin films of PBLG and Of the tao derivative PALG whose structure is
shown on the next page. The PBLG films were thoroughly
characterized by several techniques and the more fragile PALG films
by infrared spectroscopy.
114-6
Azo-modified polymer
'^NH- GH— CO- NH-
(^2)2
GO
•CH— CO^
({“2)2
COOH
114-7
III.
a. Prior to this work there was np operative spin coater iii this
laboratory. The available equipment was therefore set up and put
into working condition. All spin coating was carried out on a Solitec
5110-TC instrument. Figure 1 on the following page shows the
calibration that was carried oiit between settings and the actual
spindle speed.
Several solvents were examined for dissolution of PBLG and
PALG. For optimal results, the solid polymer must be converted to a
somewhat viscous, entirely homogeneous solution. PBLG was
dissolved successfully in 1,2-dichloroethane (DCE), dimethyl-
formamide (DMF), 1,4-dioxane (DO), and chloroform. Thin films
initially were cast on standard glass microscope slides. Both DCE and
DO gave excellent, clear, homogeneous thin films of considerable
strength. The chloroform solution was too viscous and did not
produce a uniform film. DMF spread out evenly but gave a film with
a spotty appearance. For comparison, a film also was prepared on a
glass slide by mechanical shearing with another glass slide of PBLG
dissolved in DCE. Samples also were cast on silicon or germanium
wafers. Two cycles were used on the spin coater, a slower spread
cycle and a faster spin cycle. The speed of the spread cycle typically
was 80-120 krpm for 4-10 s, and that of the spin cycle was 200-500
krpm for 20-30 sec. Actual selection of conditions determined the
thickness and quality of the film.
Attempts were made to dissolve PALG in acetone, DCE, and DO,
but homogeneous, highly colored solutions were obtained only with
DO.
114-8
Biotechnology Laboratory Spin Coater
(lAlda>l) P99ds 9|PU!ds
Spindle speed as a function of dial setting.
114-9
Figure 1
100 200 300 400 500
Dial Setting
b. Film thickness was iheasured approximately on a Dektak II
surface proHle instrument. Figure 2 on the next two pages shows
the resulting proHles from films of PBLG cast oh glass. The sample
obtained by mechanical shearing gave a rather thick and very
uneven film, With a thickness varying between 50,000 and 150,000
A, with a mean thickness of about 100,000 A (see A). The film from
DCE gave a relatively homogeneous film about 30,000 A thick (B).
That from DO was homogeneous and thinner (C). The slide from DMF
(P) gave a very spotty appearance that could not be considered a
film.
c. The films on glass were examined by polarized rhicroscopy on
a Nikon Type 109 microscope. The figure below shows the
appearance of the PBLG film out of DCE at 400x. The more viscous
material from chloroform exhibited birefringence.
Ill I
n?. • '59
'.lull |i. iii'itiiiM
05.- ir--. 9ii •'.nrro hediiim
i-- . .
- -
w.
«
•
•
«
•
r-
•
•
•
•
•: \
i L..
f\
rifi
"'1
\
— 'r—
•
1
«
; V'
-f'”
i' .
—
if
•
t
*
t
•
t
«
%
L:-
L.r.z;:.:
w,. — ..
— p.
"LRl IJ tt
riOBtr 15;i.203:..M
3S0.,fjei0
R ClIH f-i
H run n h
300.. 000
200.. 000
200.000
150.000
100. 000
50 . 000
0
.I';.;! Cif.o
I
‘'.'I.iiiih t i
Figure 2A.
Surface profile of PBLG mechanically sheared from 1,2*
dichioroethane.
1(1 i
(>•( '10 0f5 *i'(-9e
LU
■.rntl
'3PEKD.
6 . 001'
HEDll
ImI'I
.1M
' '[U-i'(
ORIE
— I..).:..,...
«
•
•
—
. .n.x.ui
«
%
«
•
«
«
’■\
IS
«
•
«
«
•
*
•
•
«
«
- -
—
. i -
—
«
«
•
V
_
u
F; CUFn 0 M e 96jl.M
M cnn H li!
50..';'>-.59 ft
1 .
60.000
50.000
40.000
30 . 000
20.000
10.000
0
■10.000
H CURSOR « 1,.673
‘.u..nrjH PEirifiK n
Figure 2B.
Surface profile of PBLG cast from 1,2-dichloroethane by spin coating.
Copy av'...a%.vj , uOiK
114-11 permit tuliy leqibk ToproductioiS
Figure 2C.
Surface profile of PBLG cast from l»4-dioxane by spin coating.
■•■f.MlI f. .OWot'l
I MSPi ofv- i<h'Sii?i sr-i-ED- rcDJiji'i
rl .
ORLT'
f; (IJR- Pi H e 9e-j>M
H ' I in n H Is!
6.- 0Ofi
SLOAIJ DEkTAK It
Figure 2D.
Surface profile of PBLG cast from N,N-dimethylformamide by spin
coating.
Copy to DTTC «o*
t
/
i
'4
I
Vi
114-12
d. In order to carry but spectroscopic studies without
interference from the glass substrate, films were cast on silicon and
germanium wafers. Samples cast from DCE or DO solutions bn silicon
were examined by spectrbscopic ellipsometry bn the Rudblph s2000
spectrometer. Figures 3 and 4 on the next two pages give the
resulting spectra. The interference pattern derives from the
thickness of the film, and other features depend on the refractive
index, as well as the thickness. These spectra are undergoing
further analysis by Capt. J. Targove of this division.
e. Infrared spectra were recorded on samples cast on
germanium or silicon with the Perkin-Elmer model 1725-X FT-IR
spectrometer. Figure 5 is of PBLG cast from DO* Features causing
rolling of the baseline are due to absence of background correction
for the germanium disc. Figure 6 is of PALG cast from DO.
The major features of the spectrum of PBLG include the NH
stretch at 3291 cm’^ and two carbonyl stretches at 1735 (ester) and
1653 (amide I) cm*^ The latter feature is indicative of an a-helix
structure.
The azo-modified derivative also has a strong NH stretch but a
more complex carbonyl region. The small peak at 1735 cm*^ is
probably from the carboxylic acid and the major peak at 1654 is the
amide I band.
114-13
t an ( Ps f )
(9QI98C9C9Q8C900
QOU^NSOtfi^ CUS
38888801^0)098
SO U) ^88 ♦ • • • •
• t # • « 18888^
- 8 8 8 8 81 1 1 I 1
X - ( B ^ t SQ ) soo
Spectroscopic ellipsometry of PBLG cast from 1,2-dichIoroethane by
spin coating.
0 id. CD CJ. 3 IQcd “SW
t an (Ps 1 )
|mmTnpnimTynmiiii]nnittTT)TnnTm}tr>niiiijiiTnnn^
CSCSQQQQcvi^r
Q 00 U) (VJ S • •
• ••••• Q Q
•x eg (S Q Q Q I I
C
nuTnitTntnniTT.iTTTTT* □
O Q (9 (Vi
U) 03 C3
f • •
Q Q -4
I 1 I
E
C
( ^ ^ I SQ ) SOO
Figure 4.
Spectroscopic ellipsome^^ PBLG cast from 1,4-dioxane by spin
coating.
Wave 1 erig'th
4000 3200 2400 2000 1600 1200 600 400
Y
IV. RECOMMENDATIONS
This work has demonstrated that homogeneous thin films of high
polymer peptides derived from polyglutamic acid may be cast by
spin coating techniques. Given a compatible substrate, these films
may be characterized by a variety of techniques, including surface
profilings polarized microscopy, spectroscopic ellipsometry, and
infrared spectroscopy. These films have a reasonably homogeneous
thickness, which can be controlled by the spin coater speed.
Consequently, their processibility properties appear to be excellent.
1
114-18
REFERENCES
Lambert. J. B., Wang, G.-t., Teramura, D. H., Interaction of the carbon-
germanium or carbon-tin bond with positive charge on a p
carbon. J. Org. Chem.. 1988, Vol. 53, pp. 5422-5428.
Levine, B. F.^ Bethea, C. G., Hyperpolarizability of a polypeptide o~
helix: poly-y-benzyl-L-glutamate, J. Chem. Phvs.. 1976, Vol. 65,
pp. 1989-1993.
Miller, R. D., Michl, J., Polysilane high polymers, Chem. Rev.. 1989,
Vol. 89, pp. 359-1410.
114-19
1990 USAF-UES SUMMER FACULTY RESEARCH PROGRAM/
GRADUATE STUDENT RESEARCH PROGRAM
Sponsored by the
AIR FORCE OFFICE OF SCIENTIFIC RESEARCH
Conducted by the
Universal Energy Systems, Inc.
FINAL REPORT
Chemical Induced Grain Boundaiy: Migration in AI3Q3
Prepared by:
Academic Rank:
Department and
University:
Research Location:
USAF Researcher:
Date:
Contract No:
Gary L Leatherman
Assistant Professor
Mechanical Engineering
Worcester Polytechnic Institute
WRDC/MLLM
Wright-Patterson AFB
Ohio 45433
Randy Hay
November 26, 1990
F49620.88-C-0053
Chemically Induced Grain Boundary
Migration in AI2O3
by
Gary L. Leatherman
ABSTRACT
Chemically induced grain boundary migration (CIGM) was observed
in bulk polycrystalline alumina. The presence of gallia coupled
with a bismuth oxide flux was able to induce the migration of
grain boundaries in the alumina. Use of the same system for
inducing grain boundary motion in alumina fibers with a "bamboo"
microstructure proved unsuccesful due to experimental
difficulties in observing the effect. Additional attempts were
made to observe CIGM in these fibers using sols of Ga203, Fe203,
and Cr203 doped with bismuth oxide. The microstructure of the
fibers prevented in the time frame of the program the conclusive
identification of CIGM.
r
115-2
I wish to thank tha Air Force Office of Scientific Research for
sponsoring this work. Universal Energy Systems was also of great
help in the non-technical aspects of the Summer Fellow Program.
115-3
I. IlifTOOpUCTIpN
Grain boundary migration which occurs in the absence of known
driving forces in which the chemical composition changes are
observed in the vplume of the crystal swept by the migrating
boundary is known as chemically induced grain boundary migration
(CIGM) ^ or as diffusion induced grain boundary migration
(DIGM)^. The driving force for thit phenomenum is believed to be
related to coherency strain energy anisotropies which develop
along the grain boundary do to the presence of solute atoms. ^
Prior to the start of this work CIGM had not been observed in
alxunihum oxide.
Alumina fibers are currently being considered as the reinforcing
element in oxide ceramic/ceramic composites. Although
polycrystalline alumina fibers are quite inexpensive, their
mechanical properties are not suitable for high performance
composite applications. These fibers, after thermal treatment
which occurs either in processing of the composite or in
service, develop a segmented ('•bamboo”) grain structure. Once
this structure is obtained there is no further driving force for
conventional grain growth. The grain boundaries are quite weak;
failure originates at these points. This limits the effective
length of the fiber to the "bamboo” segment length which is not
long enough to provide effective reinforcement to the matrix.
Single crystal saphire fibers overcome this problem. However,
they are extremely expensive. Current costs are of the order of
hundreds of thousands of dollars per kilogram.
115-4
II. OBJECTIVES OF THE RESEARCH EFFORT
The development of a polycrystalline aliuniha. fiber with
sufficiently long bamboo segment length could be a low cost
alternative to single crystal saphire fiber. CIGM may provide a
mechanism by which further grain growth can be occur in alumina
fiber. Thus providing a suitable segment length for the fibers
resulting in adequate composite properties. This project
examines CIGM in alumina. The first objective was to observe
CIGM in bulk polycrystalline alumina. After the presence of CIGM
in bulk alumina was established, the next step was to determine
if CIGM occurs in bamboo structure alumina fiber.
III. RESULTS OF RESEARCH EFFORT
Specimens of polycrystalline alumina were placed in a bed of
gallia powder doped with 10 weight percent Bi203. The specimens
were heated to 1320 'C and held for 24 hours. Tapered sections of
the surface exposed to the powder bed were porduced using
standard petrographic techniques. Figure 1 shows a typical
section of an alumina specimen which underwent the above thermal
treatment. It is typical of a CIGM microstructure. This is
strong evidence that CIGM does occur in alumina. In addition,
during the course of this work researchers in Korea published
results also showing that CIGM occurs in alumina^ . Thus the
first objective of the program was reached.
115-5
Attempts to observe CIGM in alumina fibers using an identical
approach were uhsuccesful. The bismuth oxide flux created a film
on the fibers that prevented observation of the microstructiire .
An alternative approach was attempted using sol-gel techniques.
Sols based on Ga203, Fe203, and Cr203 were developed. Alumina
fibers were then coated with these sols and thermally treated at
1320 *c for 24 hours. Although the microstructure of these fibers
could be obsexrved under polarized light in transmission, direct
evidence of net grain growth do to CIGM was not obtained. The
large distribution in the grain size of the fibers before CIGM
treatment made any before and after observations difficult to
interpret without detailed statistical analysis of both the
before and after grain size distributions. There was not
sufficient time in the program to do this sort of analysis.
Longitudinal cross sections of the fibers were examined by
scanning electron microscopy (SEM) for compositional gradients
typical of CIGM. Difficulties in producing quality samples as
well as equipment problems and the time limit of the program led
to inconclusive results.
IV. RECOMMENDATIONS
Although there is conclusive evidence for CIGM in alumina, there
is not yet sufficient information on CIGM's presence or extent
in alumina fibers. The following work should be done to answer
this question. First, detailed grain counting for complete
characterization of the grain size distribution in alumina
fibers should be conducted before and after CIGM exposure. This
should determine if on average there is any increase in bamboo
115-6
segment length due to grain boundary migration. Second,
longitudinal cross section of fibers that have been exposed to
the CIGM environment at a variety of depths should be produced
and carefully examined using an electron microprobe. This should
detemine if CIGM is occuring in alumina fibers and if it is a
useful process by which to induce grain growth in alumina
fibers.
115-7
REFERENCES
1. Hillert, M. and Purdy, G.R., "Chemically Induced Grain
Boundary Migration," Acta Met. 1978, Vol, 26, pp. 333-340.
2. Balluffi, R.W. and Cahn, J.W. , "Mechanism for Diffusion
Induced Grain Boundary Migration," Acta Met. 1981, Vol. 29,
pp. 493-500.
3. Hillert, M. , "Oh the Driving Force for Diffusion Induced Grain
Boundary Migration," Scrinta Met. 1983, Vol. 17, pp. 237-240.
4. Lee, H. and Kang, S., "Chemically Induced Grain Boundary
Migration and Recrystallization in AI2O3," Acta Met. 1990,
Vol. 39, pp. 1307-1312.
115-8
Figure 1: S.E.M. image and 0, Al, and Ga E.D.S. maps
of polycrystalline alumina sample exhibiting
gallia driven chemically induced grain
boundary migration.
115-9
1990 USAF-UES SUMMER FACULTY RESEARCH PROGRAM/
GRADUATE STUDENT RESEARCH PROGRAM
Sponsored by the
AIR FORCE OFnCE OF SCffiNTMC RESEARCH
Conducted by the
Universal Energy Systems, Inc.
FINAL REPORT
On the Use of OPA (Qualitative Process Automation! for Batch Reactor O
Prepared by:
Academic Rank:
Department and
University:
Research Location;
USAF Researcher:
Won-Kyoo Lee, Ph.D.
Associate Professor
Department of Chemical Engineering
The Ohio State University
WRDC/MLIM
Wright-Patterson AFB
WPAFB,OH 45433-6533
Maj Steven R. LeClair
21 Aug90
F49620-88-C-0053
Date:
Contract No;
On the Use of QPA (Qualitative Process Automation^ for Batch Reactor Control
by
Won-Kyoo Lee
ABSIRACr
Control of batch reactors and the selMirected process control system, QPA, were reviewed
to determine if the QPA system could be used for intelligent control of batch reactors. The
control of batch reactors has been formulated as optimal control problems, with the solution
being an openJodp temperature trajectory. However, this optimal temperature profile is
based on Very complicated, but still incomplete mathematical models to account for the
unique nonlinear and time-varying dynamics of batch reactors. This means that the greatest
remaining challenge in controlling batch reactors is to develop a totally adaptive control
strategy that can result in the optimal operation using a minimum of mathematical models.
In this regard, the capability of the QPA control is expected to be more beneficial for batch
reactors, especially in the presence of process changes, and the dynamic, nonlinear nature
of the batch reactors. It is suggested that the QPA system be tested to further demonstrate
its concept and consequently extend its applicability by being applied to an experimental
unit.
116-2
ACKNOWLEDGEMENTS
I ■ i^h to thank the Air Force Systems Command and The Air Force Office of Scientific
Research for sponsorship of this research. Universal Energy Systems must be mentioned
for their concern and help to me in all administrative and directional aspects of this
program.
My experience was eye-opener, and subsequently I was able to broaden my background by
exposing myself to on-going manufacturing research projects. Maj Steve LeClair provided
me with support, and enjoyable research environment.
116-3
1. INTRODUCTION:
Batch processes have gained their importance in recent years as many chemical industries
seek the manufacture of low- volume, high-value-added chemicals. In particular, batch
reactors are extensively used in producing specialty chetnicals, polymers, new drugs, etc.,
because of their great flexibility and rapid response to changing muket conditions. For
batch reactors, the high valued final product makes determination of an optimal control
profile important, especially for maximization of some function of the composition (e.g.,
productivity, yield, or selectivity), these problems have been formulated as optimal
control problems, with the solution being an open-loop temperature or flow-rate trajectory.
However, these optimal control profiles are difficult to obtain because of the absence or
excessive development cost of adequate models. Another factor that complicates the
determination of these optimal control profiles is the presence of constraints on both the
control and state profiles.
The Manufacturing Research Group of the Materials Laboratory at Wright Research and
Development Center, Wright-Patterson Air Force Base, is concerned with the development
of intelligent (self-directing and self-improving) systems for real-time control of advanced
material processes. As a result, a self-directed process control philosophy, referred to as
QPA (Qualitative Process Automation), was developed for the control of autoclave curing
process of composite materials by making use of, not only qualitative physics, but also
concurrent expert systems cooperating together. More specifically, the QPA system is
capable of automating, not only the control of the process, but also the development of an
optimized cycle. Further, because QPA develops the processing knowledge adaptively in
situ, this capability is expected to be more beneficial for batch chemical reactors in the
presence of process changes and the dynamic, nonlinear behavior of the batch reactors.
1164
My research intdcsts have been in the areas of adaptive control, on-line optimizing control,
model-based control, statistical process control, neural networks, genetic algorithm and
expert systeni framework for process control system synthesis, with a balance between
theory and application. My work on adaptive strategies for the automatic startup and
control of batch processes contributed to my assignment to the Manufacturing Research
Group of the Materials Laboratory.
116-5
n. OBJECTIVES OF THE RESEARCH EFFORT;
The increasing interest in the manufacture of low-volume specialty chemicds in batch
reactors will continue to present many process control challenges. First, the frequent
product and process changes associated with a batch chemical manufacturing facility
requires an optimal startup policy for smooth and rapid transition to the desired operating
conditions with little or no overshoot. The swond type of problem is to control batch
reactors to optimize composition functions such as productivity. These problems have
been formulated as optimal control problems, with the solution being an open-loop
temperature or flow-rate trajectory. Techniques are available to solve for the optimal
control profiles, once good process and kinetics models are known. However, these
optimal solutions are based on very complicated, but still incomplete mathematical models
to account for the unique nonlinear and time-varying dynamics of batch processes. This
means that in controlling batch reactCH^, often the greatest challenge is the development of a
totally adaptive control strategy that requires a minimum of mathematical models and
consequently can be applied to many batch processes.
My assignment as a participant in the 1990 Summer Faculty Research Program (SFRP)
was to examine the attributes specific to batch reactor control and then determine if the self-
directed process control system, referred to as QPA, researched and developed at the
Materials Laboratory could be used for intelligent control of batch reactors.
116-6
m. BATCH REACTOR CONTROL:
In general batch reactors are characterized by both very different manufacturing
environment and different dynamic behavior. For example, the multiproduct environment
of batch reactors requires that ingredients, control loop setpoints, and tuning parameters
must be changed frequently. The frequent product and process changes demand good
dynamic response over the entire operating range of the controlled variable for startup and
shutdown regulatory control. Also, the wide operating ranges and nonstadonary behavior
that cause difficult sensor problems in batch reactors influence control system design. The
controller design is further complicated by asymmetric penalties, such as in composition
control where the formation of unwanted byproducts is irreversible. In the conux)! of batch
reactors, there are two related but distinct challenges.
(a) The first is the control of batch reactors with highly exothermic reactions. Safety-
related concerns such as adiabatic runaway are primary, followed by productivity concern
if the heat-release rate is far from constant.
(b) The second control challenge is that of controlling the batch reactor to optimize a
composition function such as yield or selectivity.
There has been considerable interest in the past on the control of batch reactors. However,
most of the previous work has focused on either the determination of the optimal
temperature trajectory based on reaction kinetics alone or on the design of control systems
for tracking this predetermined temperature profile. Also, in practice the operation of a
batch reactor is formulated in terms of a temperature trajectory because temperature is the
most readily available output. The trajectory is specified as a sequence of steps, consisting
of three parts: (a) startup, where the reactor contents are brought from the initial charging
conditions to the desired operating level; (b) maintenance of the desired nominal operating
conditions for as long as it is beneficial to do so; (c) termination of the reaction according to
116-7
dther optimality or product speciHcations considerations. Thus, the discussion ofbatch
reactor control is divided into three parts. The first discusses the automatic sta^p and
isothermal control Of batch reactors. The second and third discuss the optimal control of
batch reactors to m^imize some function of ctmiposition.
A. Automatic Startup and Isothcmial Batch Rgactor Conirol
For exothermic batch reactors that undergo frequent startups and shutdowns, the startup
phase constitutes an important part of the batch cycle in enhancing productivity. It is,
therefore, most often desirable to provide a smooth and rapid transition to the desired
operating condition with little or no overshoot. This can be achieved by applying time-
optimal control as an optimal startup policy. The time-optimal control involves switching
of the control input from one constraint to the other at predetermined times. For example,
the control ihput needs to be switched only once to achieve the most rapid transition for
second-order, single-input/single-output systems, and the switching time can be calculated
a priori or online from a second-order process model. Alternatively, PID controlli?^ can
also be used to follow a pre-determined time-optimal trajectory.
However, both time-optimal control with predetermined switching time and PDD controller
have difficulty in accommodating changing process characteristics that could be
experienced from batch to batch or product to product. These process changes and the
dynamic, nonlinear nature of batch reactors require recalculation of the switching times, the
time-optimal trajectory or new settings of PID controller for satisfactory reb> ;lts. Moreover,
for batch reactors with unknown or poorly understood dynamics, neither of these
techniques can provide satisfactory control. To alleviate the problems with varying process
dynamics, a model-predictive time-optimal control method and an adaptive time-optimal
control algorithm were proposed for the startup control problem, respectively. Another
116-8
adaptive control techniques using a linear dynamic model have been applied to the startup
problems for batch reactors since these adaptive control schemes provide systematic,
flexible approaches for dealing with uncertainties, nonlineaiities, and tiine-varying process
parameters. In particular, the self-tuning controller (STC) of Clarke and Gawthrop was
applied to control of batch reactors for the mtuiufacture of synthetic latex of a variety of
grades. The general control objective was to produce polymer of acceptable quality in the
shortest practicable time. For this purpose, the batch temperature setpoint was
predetermined for each grade, and a control scheme was used to follow the predetermined
batch temperature profile as close as possible without overshoot
Recently, an adaptive strategy based on the prediction error model was proposed to achieve
desired temperature profiles during startup of a continuous fluidized sand bath, and then
later extended to the automatic stturtup and control of a batch heating process and a
simulated exothermic batch reactor. Prediction error models contain information about the
quality of the prediction, and thus provide better predictions of the future output variable,
which is used for on-line calculation of the startup switching time. For regulation after the
setpoint is reached, a self-tuning controller with the PID structure was applied to obtaining
PID controller settings, and a long-range predictive control strategy was also proposed by
incorporating the prediction error model in an extended horizon approach. Results have
shown the feasibility of the adaptive approach based oh prediction error models for the
rapid startup and control of exothermic batch reactors. However, there are limitations and
potential operating problems associated with current adaptive control schemes. The biggest
drawback of adaptive control techniques is the inherent linear nature of an assumed
dynamic model. Moreover, the range of uncertainty may be substantially greater than can
be tolerated by existing algorithms for adaptive systems.
H
116-9
B. ^Qrii5otliermal.Batch .Rcagtor Control
Although the previous discussion of regulatory temperature control has focused on startup
and isothermal reactor operations, nbnisothermal temperature trajectories are a common
means of optimizing the throughput of many batch reactors. This is because the
productivity is limited by the maximum heat-generadon rate, and this typically occurs for
only a short time. For exothermic reactions, the temperature trajectories are often calculated
to maintain a constant reaction rate, which is within a specified safety limit of the maximum
heat-removal capacity of the reactor. Determination of the constant-rate temperature
trajectories requires a mathematical rate expression, and funher the solution of vmational
optimization problem or experimental determination is required when rate equations are
complex. These temperature profiles are then implemented as a series of open-loop
setpoint changes in the reactor temperature controller. A major.<problcm with this open-
loop strategy is that induction effects in the chemical reactions introduce uncertainties in the
timing of the setpoint changes.
Alternative approach to maintaining constant-rate operation in the face of unknown or
uncertain kinetic is to obtain a real-time measurement of the reaction rate. However, on¬
line analytical instruments that can measure properties of interest are often not available.
Since almost all reactors are or can be well instrumented with temperature and flow
measurements, these measurements can be combined with a set of dynamic energy balance
equations for the reactor in order to monitor in real time the instantaneous heat release due
to reaction. Form this the instantaneous rate of reaction can be evaluated, and subsequently
one can obtain an estimate of the conversion of material in the reactor at any instant of time.
These quantities are then very useful in monitoring the reactor for potentially dangerous
conditions, in improving temperature control, and in improving the control of composition
and other quality variables.
116-10
The objective here is to manipulate the temperature in order to optimire some function of
the composition (e.g., productivity, yield, or selectivity). These problems have been
formulate as optimal control problems, with the solution being an qren-loop temperature
trajectory. Techniques are in hand to solve for the optimal temperature trajectory once good
process and kinetic models are known. However, these optimal control profiles are difficult
to obtain because of the absence or excessive development cost of adequate mathematical
models, especially for cases where the reaction kinetics and mechanisms are unknown or
poorly known. Again, a major problem with implementing this open-loop strategy is that
variations in the process variables and parameters during each cycle and/or from cycle to
cycle require recalculation of the optimal control trajectory, as with the control of
nonisothermal batch reactors.
IV. JUSTIFCATION OF QPA FOR USE IN BATCH REACTOR CONTROL:
As discussed previously, the optimal operation of batch reactors has been formulated as
optimal control problems, with the solution being an optimal temperature profile based on
mathematical models. And then, control systems have been designed to guide the batch
reactors along this trajectory. There is a major problem with this open-loop strategy in that
mathematical solutions are limited in accommodating changing process conditions and/or
variables which can not be accounted for. It is thus clear that there is a great need to
develop an on-line control algorithm that adaptively results in optimum operation in the
presence of process changes and the dynamic, nonlinear nature of batch reactors, by using
a minimum of mathematical models.
116-11
QPA is ah intelligent process condrol system based on “goal-driven” control philosophy,
which is capable of developing the processing knowledge adaptively in ritu, generating an
optimized cycle, and automating the control of the process to achieve the desired product
goal. These capabilities are made possible by the use of qualitative physics because of the
reduced computational load on the control system and the ability to accommodate both
uncertain and nonlinear process characteristics through qualitative relationships. More
specifically, the QPA system autonomously interprets heterogeneous sensor data and
represents that data qualitatively using process-specific knowledge coupled with expert
heuristics, and reasons about the data to construct a processing plan in real time. In
summary, QPA offers the capability which are c(»isidered to be desired features of a totally
adaptive control system for batch reactors.
A prototype QPA system for autoclave cupe of graphite-epoxy composites has been
demonstrated to successfully control the autoclave process for making advanced composite
materials. In addition, this QPA system has demonstrated dramatic improvements in
material processing times as a result of its ability to adapt on-line to material processing
behavior and to generate tailored cure cycles. It should be noted that the autoclave is a
batch reactor which requires a startup period for heating of the matrix resin to start flow and
chemical reaction, followed by a cure period for completing of the reaction and a cooldown
period for removing of the cured laminate. The control objective was to complete the
process as rapidly as possible without damage by the potentially excessive heat generated
by the reaction itself. As described above, the autoclave curing of graphite-epoxy exhibits
the same process characteristics and thus has the same control problem domain issues as
addressed in controlling the exothemic batch reactors.
116-12
V. RECOMMENDATIONS:
Real-time control of batch reactors is difficult by conventional control thetny because of the
absence or excessive development cost of adequate mathematical models to account for
uncertainties, nbnlinearities and time-varying process parameters. Another factor that
complicates the use of conventional control theory is the existence of transitions between
processing stages with different kinds of phenomena and physical objectives. The
capability of the QPA system and its proven application to the autoclave curing of
composite materials being considered, control of batch reactors is expected to greatly
benefit from QPA.
A follow-on research is suggested on the use of QPA for the control of batch reactors to
further demonstrate the concept of QPA and extend its applications. The QPA will be
tested by being applied to an experimental batch reactor, possibly a batch polymerization
reactor, under the sponsorship of an.AFOSR Mini Grant. To this end, the QPA software
will be obtained from Universal Technology Corporation, Dayton, Ohio, and effort will be
made to kek an industrial partner.
REFERENCES
Clarice, D.W., and PJ. Gawthrop. Self-Tuning Controller. Proc. lEE Part D. 1975, Vol.
122, pp. 929-934.
Clarke, D.W., and P.J. Gawthrop. Self-Tuning Control. Proc. lEE Part D. 1979, Vol.
126, pp. 633-640.
Cuthiell, J.E., and L.T. Biegler. Simultaneous Optimization and Solution Methods for
Batch Reactor Control Profiles. Computers Chem. Engng. 1989, Vol. 13, pp. 49-62.
De Keyser, R.M.G., and A.R. Van Cauwenberghe. A Self-Tuning Multistep Predictor
Application. Automatica. 1981, Vol. 17, pp. 167-174.
Hodgson, A.J.F., and D.W. Clarke. Self-Tuning Applied to Batch Reactors. Proc. lEE
Vocation Sch. ,Ifld.J?igilal Control System. i£E. London, 1984.
Hsu, E.H., S. Bacher, and A. Kaufman. A Self-Adapting Time-Optimal Control
Algorithm for Second-Order Processes. AIChE J. 1972, Vol. 18, pp. 133-139.
Juba, M.R., and J.W. Hamer. Progress and Challenges in Batch Prcxiess Control. Proc.
of the Third International Conference on Chemical Process control. Asilomar, CA, 1986.
Koppel, L.B., and L.R. Latour. Time Optimal Control of Second-Order Overdamped
Systems with Transportation Lag. Ind. Engng. Chem. Fundam. 1965, Vol. 4, pp. 463-
471.
116-14
Latour, P.R., L.B. Koppel^ and D.R. Goughanowr. Time-Optimum Control of Chemical
Porocesses for Set Point Changes. Ind. Eneng. Chem. Process Des. Dev. 1967, Vol. 6,
pp. 452-460.
Latour, P.R., L.B. Koppel, and D.R. Goughanowr. Feedback Time-Optimum Process
Controllers. Ind. Engne. Chem. Process Des. Dev. 1968, Vol. 7, pp. 345-353.
I^Clair, S.R., and F.L. Abrams. Qualitative Process Automation.
Integrated Manufacturing. 1989, Vol. 2, pp. 205-211.
LeClair, S.R., F.L. Abrams, and R.F. Matejka. Qualitative Process Automation: Self-
Directed Manufacture of Composite Materials. AI EDAM. 1989, Vol. 3, pp. 125-136
Lewin, D.R.^ and R. Lavie. Designing and Implementing Trajectories in an Exothermic
Batch Chemical Reactor. Ind. Eng. Chem. Res. 1990, Vol. 29, pp. 89-96.
MacGregor, J.F. On-Line Reactor Energy Balances via Kalman Filtering.
ries. Akron, OH, October 1986, pp. 27-31.
Manzini, R.A., and E.A. Roehl. Flexible Control of an Organic Matrix Composite Cure
Process Using Object-Oriented Control Concepts. Proc. 1990 American Control
Conference. San Diego, CA, May 1990, pp. 1980-1985.
Mellichamp, D.A. Model Predictive Time-Optimal Control of Second-Order Processes.
Ind. Eng. Chem. Process Des, Dev. 1970, Vol. 9, pp. 494.
116-15
Merkle, J.E., and Won-Kyoo Lee. Adaptive Strategies For Automatic Startup and
Teniperature Control of a Batch Process. Computers Chem. Engng. 1989, Vol. 13, pp.
87-103.
Ozkan, A.T., and Won-Kyoo Lee. Adaptive Strategies Applied to Autorriatic Startup and
Control of an Exothermic Batch Reactor. Proc. 1989 American Control Conference.
Pittsburge, PA, June 1989, pp. 1834-1839.
Pardee, W.J., M.A. Shaff, and B. Hayes-Roth. Intelligent Control of Complex Materials
Processes. AI EDAM. 1990, Vol. 4, pp. 55-65.
Phillips, S., D.E. Seborg, and K.J. Legal. Adaptive Control Strategies or Achieving
Desired Temperature Profiles during Process Startup. 1. Model Development and
Simulation Studies. Ind. Eng. Chem. Research. 1988, Vol. 27, pp. 1434-1443.
Phillips, S., D.E. Seborg, and K.J. Legal. Adaptive Control Strategies for Achieving
Temperature Profiles during Start-up. 2. Experimental Application. Ind. Eng. Chem.
Research. 1988, Vol. 27, pp. 1444.
116-16
1990 USAF-UES SUMMER FACULTY RESEARCH PROGRAM/
GRADUATE STUDENT RESEARCH PROGRAM
sponsored by the
AIR FORCE OFFICE OF SCIENTIFIC RESEARCH
conducted by the
Universal Energy Systems, Inc.
FINAL. R££fiSI
ULTRASONIC TECHNIQUES FOR AUTOMATED DETECTION OF
FATIGUE MICROCRACK INITIATION AND OPENING BEHAVIOR
Michael T. Resch, Ph.D.
Assistant Professor
Engineering Mechanics
University of Nebraska* Lincoln
WRDC/MLLN
Building 655, Room 023
Wright- Patterson AF&, OH 45433
Theodore Nicholas, Ph.D.
Date:
Prepared by:
Academic Rank:
Department and
University:
Research Location:
USAF Researcher:
Contract No.:
September 30, 1990.
F49620-88-C-0053
UI.TRASONIC TECHNIQUES FOR AUTOMATED DETECTION OF
FATIGUE MICROCRACK INITIATION AND OPENING BEHAVIOR
by
Michael T. Reach, Ph.D,
ABSTRACT
A surface acousclc wave non* destructive evaluation technique was used to
detect the natural nucleatlon of surface microcracks In highly stressed
regions of hourglass shaped aluminum specimens during fatigue cycling.
The experimental procedure Involved excitation of Rayleigh waves on the
surface of each specimen and observation of the presence of a specular
reflection from the nucleating crack superimposed on nonspecular
reflections from mlcrostructural features surrounding the flaw.
Contacting wedge transducers were used to excite the incident waves and
to detect the reflected wave signals. The effectiveness of a split*
spectrum processing algorithm to improve the minimum detectable crack
size of isolated cracks in the scattering field was demonstrated.
Additionally, measurements of crack opening behavior were performed both
acous.tically and with the laser Interference displacement gage. Initial
results indicate that the acoustic technique is more sensitive to small
traction forces oh adjacent crack faces than is the laser interference
technique.
\\r-2
Acknowledgements
I wish to thank the Air Force Systems Command and the Air Force Office
of Scientific Research for sponsorship of this research. Universal
Energy Systems must be mentioned for their concern and help to me in all
administrative and directional aspects of this program.
My experience was rewarding and enriching because of many different
influences. Both Ted Nicholas and Jay Jira of the high temperature
metals and ceramics group went out of their way to ensure that I was
supplied with all the instrumentation and technical help necessary to
address my stated research goals and objectives. Special thanks are due
to Prasanna Karpur for sharing his expertise in signal processing of the
acquired signals. All split* spectrum processing of the data obtained in
these experiments was performed courtesy of Dr. Karpur utilizing his
personal software. Finally, the comeraderie and experimental expertise
of Rick Kleisraet of U.D.R.I. was invaluable in the preparation of
fatigue specimens and operation of the laser interference displacement
gage equipment.
117-3
I. INTRODUCTION:
The reason for developing new quantitative nondestructive evaluation
techniques to measure the size and opening behavior for surface
microcracks is that in the so called small crack size regime small
cracks have been observed to grow at rates which are orders of magnitude
higher than large sized cracks when subjected to identical magnitudes of
crack driving force. Quantitative measurements of crack depth below the
surface for surface uicrocracks facilitate the evaluation of crack
growth rate V; the number of cycles. Nondestructive measurements of
crack opening behavior are especially important here because many
current theories which address the issue of why small cracks grow faster
postulate that small cracks have less closure . than large cracks,
resulting in a higher driving force for growth for small cracks.
The High Temperature Metals and Ceramics Branch of the Uright Research
and Development Center is particularly concerned with the creation of
practical techniques to facilitate automation of fatigue testing of
materials important to the mission of the Air Force both under
laboratory conditions and on structures operating in the field. The
expertise of this branch in the development of innovative and useful
techniques to characterize the kinetics of growth of microscopic surface
fatigue cracks in metals, ceramics, and composites is well documented in
the international literature.
My research interests have been in the area of development of surface
1174
acoustic wave scattering techniques to measure the size, growth rate,
and opening behavior of small, surface cracks [1,2], Therefore, a
project concerning evaluation of the feasiblity of using this technique
in automated fatigue crack experiments seems a logical extension of the
technology.
II. OBJECTIVES OF THE RESEARCH EFFORT:
Currently, nondejtructive detection of initiation of surface microcracks
in hourglass shaped specimens now in use for fatigue crack initiation
studies in the High Temperature Metals and Ceramics Branch of the Wright
Research and Development Center (WRDC) is accomplished by optical
scanning ac high magnification in a standard metallographic microscope.
This process is exceedingly time consuming, requiring periodic
inspections of the specimen performed by physically taking the specimen
out of the servohydraulic machine, scanning many square millimeters of
surface ar«a in the high stress region, and replacing and realigning the
specimen ii the hydraulically actuated grips before resuming the test.
Additionally, during crack mouth opening displacement measurements of
a]ready nucleated fatigue microcracks with a Laser Interference
Displacement Gage (LIDG) , there is uncertainty concerning the exact
amount of tensile stress to be applied to the specimen to cause the
adjacent crack faces of the microcrack to become traction free.
Surface Acoustic Wave (SAW) scattering has the potential to contribute
to this effort in two ways. First, a beam of surface acoustic waves
117-5
directed toward the high stress region of a specimen interrogates the
total area on one side of the specimen where microcrack initiation
technique is expected to occur. Measurement of the reflection of the
waves is accomplished in a few seconds during a single measurement
without removing the specimen from the testing machine. Surface
microcracks as small as 50 micrometers deep have been detected using
this technique [1,2]. Second, measurement of the reflected amplitude of
surface acoustic waves from a surface microcrack as a function of the
amplitude of applied tensile stress normal to the plane of the adjacent
crack faces has demonstrated that ultrasonic measurements of opening
behavior correlate well with measurements of Crack Mouth Opening
Displacement (CMOD) and Crack Tip Opening Displacement (CTOD) performed
under stress in a scanning electron microscope [2].
The chief physical phenomenon which stands as an obstacle to the
utilization of the SAW scattering technique in these experiments is that
the metallurgical features in complex alloys contribute to the
scattered signal which also contains the reflected echo from a
nucleating fatigue crack. The primary goal has been to acquire
scattered signals from the high stress region of hourglass shaped
fatigue specimens at periodic intervals, and examine the acquired
waveforms for evidence of earliest possible detectability of reflections
from surface microcracks when the amplitude of the reflection is
obscured by microstructural scattering. The secondary goal is to
compare measurements of crack opening behavior using the SAW scattering
technique with measurements obtained using the LIDG technique.
117-6
III. NEW HOURGLASS SHAPED SPECIMENS
a. Design and fabricate new hourglass shaped specimens similar to
specimens now in use in the high temperature metals and ceramics
laboratory which are suitable for; electropolishing, gripping by self
aligning hydraulic grips in use at WRDC, and which allow use of new
«
surface acoustic wave wedge transducers developed by this author at the
University of Nebraska-Lincoln.
b. Specimens were jproduced with a configuration used by Larsen et
al.[4], except with a major cross section of 12.5 mm by 12.5 mm, and a
reduced section of 3 mm by 12.5 mm, giving a geometrical stress
concentration factor of 1.2 . Additionally, the specimens are initially
machined with 37.5 mm by 50 mm rectangular ends suitable for insertion
in standard compact tension fracture mechanics clevis grips using 12.5
mm diameter pins. This allows fatigue initiation experiments to be
conducted at locations which do not yet possess self aligning hydraulic
grips. The enlarged areas may be cut off subsequent to fatigue
initiation experiments to facilitate insertion in hydrauic testing
machines which feature a laser interference displacement gage for
additional measurements concerning crack opening behavior.
IV. SWITCHING SYSTEM FOR TWO DUAL- ELEMENT ARRAYS
a. Build switching box to accomodate two dual-element transducers to
enable alternate surface acoustic wave scanning of both sides of the
high stress areas on hourglass shaped specimens during fatigue cycling.
b. A switch box was designed and fabricated to allow manual switching
between two sets of dual element surface acoustic wave contacting wedge
transducers using a double pole, double throw manual switch. This
enabled both sides of a fatigue specimen to be interrogated at periodic
intervals during fatigue cycling without disturbing the contact between
the transducers and the specimen.
V. NATURAL INITIATION OF FATIGUE MICROCRACKS
a. Produce naturally initiated cracks during fatigue cycling on
hourglass specimens of aluminum using the SAW scattering technique as
the primary feedback mechanism for detecting microcrack initiation.
b. An hourglass shaped specimen is metallographically prepared using
conventional wet grinding, diamond polishing, and electropolishing
techniques in the high stress region to minimize the effect of
fabrication on surface roughness and residual stress. The signal due to
backscattered Rayleigh waves from the high stress region is acquired
with a digitizing oscilloscope during application of a tensile stress of
275 MPa . Then at 2000 cycle intervals of applied fluctuating stress
(maximum stress equal to 275 MPa with a stress ratio of 0.1) this
measurement is repeated. This procedure continues until at least one
crack is produced in the high stress region which just exceeds the
maximum amplitude of the microstructurally induced interference pattern.
117-8
A crack is relatively easy to detect in the interference pattern because
its reflection Is only visible with an applied tensile stress applied to
the specimen which is large enough to completely separate adjacent crack
faces .
VI. WAVEFORM TRANSFER SOFTWARE
a. Obtain digitized waveforms from crack initiation area at regular
intervals during cycling and transfer them to the hard disk of an IBM
FC'AT for subsequent transfer to the VAX system using PROCOMM software.
b. A program was written in BASICA language to transfer the signal
obtained from a Hewlett Packard 54201A digitizing oscilloscope to the
screen of a CGA monitor and write the 1001 data points in the signal to
the hard disk utilizing a HP-IB card designed for the IBM PC-AT. Each
data file was then transferred to the VAX computer system for subsequent
split spectrum processing.
VII. SPLIT SPECTRUM PROCESSING
a. Use Split Spectrum Processing (SSP, a nonlinear digital signal
processing technique as implemented by Dr. P. Karpur of UDRI) to
determine the absolute minimum detectable size and the number of cycles
at which this size is detected [5,6].
117-9
b. After a single crack is detected in the scattered field,- the
specimen is removed from the testing machine and examined at high
magnification in a metallographic microscope. The length of the crack
at the surface is measured using a calibrated eyepiece. This single
measurement of crack size is then used to obtain the scattering
parameters necessary for scaling the size of the crack from ultrasonic
scattering theory [1]. Split spectrum processing is performed on the
set of all waveforms obtained during periodic examination of the
specimen to remove the nonspecular reflections of micros true tural
origin. This process reveals the size of the specular reflections from
the crack obtained from each cycling interval during the .experiment. The
minimum detectable crack size for a naturally Initiated surface
microcrack in 2024 aluminum using this technique was determined to be 20
micrometers [3]. This is substantially smaller than the
microstructurally llisited minimum detectable crack size of 80
micrometers previously reported in the literature for this material [1].
VIII. SAW NDE SOFTWARE FOR OPENING MEASUREMENTS
a. Develop software for computerized data acquisition of Surface
Acoustic Wave NonDestructive Evaluation (SAW NDE) information concerning
the amplitude of the reflected echo from the crack as a function of
applied stress during application of stress to the specimen utilizing a
servohydraulic testing machine.
117-10
b. A program was written in BASICA language to transfer the amplitude
of a time gated signal containing the reflection from an isolated
surface microcrack obtained from a Hewlett Packard 54201A digitizing
oscilloscope utilizing a HP- IB card designed for the IBM PC- AT,
Additionally, a DATA TRANSLATION 2818 analog to digital conversion
system was used to detect the signal from the load cell of the
servohydraulic testing machine. Under manual control of the set point
of the servohydraulic system in load control, applied force v. crack
amplitude could' be acquired automatically point by point, and plotted
graphically on the CGA screen of the PC-AT, The results of a typical
scan from an Isolated surface microcrack are displayed in Figure 1. The
nonlinear portion of the curve from zero force up to the point at which
the plot becomes vertical reveals the presence of surface tractions
between adjacent crack faces. The vertical portion reveals the point at
which surface tractions disappear, and the crack is fully open,
IX, SAW NDE V, LIDG MEASUREMENTS OF CRACK OPENING
a. Obtain force v. CMOD for naturally nucleated surface raicrocracks in
specimens of aluminum using the Laser Interference Displacement Gage
(LIDG) at identical levels of force as the ultrasonic measurments and
compare the experimental results,
b. Force v, CMOD for several naturally initiated surface raicrocracks
was accomplished using the LIDG apparatus available on machine #1 in the
high temperature metals and ceramics laboratory at WRDC, A typical
117-11
lS.8
13.5
12.Q
Figure 1
Plot of crack reflection aaplitude v. force applied normal to
the crack face obtained from the SAW NDE scattering technique
Figure 2. Plot cf crick mouth opening displacement v. force applied
norroil to the crack face obtained from the LiDG technique.
result of this type of measurement obtained for the crack which was also
ultrasonically measured is shown in Figure 2. An important feature is
the observed linear relationship between CMOD and applied stress which
is distinctly nonlinear in the ultrasonic measurement of opening
behavior of this crack shown in Figure 1,
X. RECOMMENDATIONS:
a. Periodic measurements of surface acoustic wave scattering from the
high stress region of fatigue specimens have demonstrated that the split
spectrum processing technique is capable of significantly decreasing the
minimum detectable crack size of naturally initiated surface
microcracks. Comparison of ultrasonic measurements of naturally
initiated surface micfocracks with laser measurements reveals that the
ultrasonic technique appears to be more sensitive to the existence of
small tractions on adjacent crack faces than the laser technique.
Consequently, these results demonstrate that the surface acoustic wave
scattering technique provides useful information during automated
microcrack initiation experiments, and may be easily integrated into the
testing program of the high temperature metals and ceramics laboratory
for further research and development.
b. Development of the technique is limited by the lack of contacting
wedge transducers small enough for a dual element array to fit on the
standard sized 7.5 mm wide specimens used in the high temperature metals
and ceramics laboratory. Consequently, design and fabrication of new
117-13
miniature dual element contacting wedge transducer arrays is necessary
for implementation of the technique during automated detection of
microcrack initiation on standard sized specimens. This will involve
experimental research on the physics of wave generation in commercially
available piezoelectric ceramics in order to characterize the maximum
allowable power levels which can be achieved, in addition to the
application of the principles of physical acoustics in order to
establish the optimum level of miniaturization possible using
available materials.
REFERENCES
1 Resch, M.T., Nelson, D.V. , Yuce, H.H., and Ramusat, G.F. , "A
Surface Acoustic Wave Technique for Monitoring the Growth Behavior
of Small Surface Fatigue Cracks", Journal of Nondestructive
Evaluation. Vol, 5, No. 1, 1985, pp. 1-7.
2 Resch, M.T. , Nelson, D.V., Yuce, H.H. and London, B.D., "Use of
Nondestructive Evaluation Techniques in Studies of Small Fatigue
Cracks", Basic Questions In Fatigue: Volume I. ASTM STP 924, J.T.
Fong and R.J. Fields, eds., American Society for Testing and
Materials, Philadelphia, 1988, pp. 323-336.
117-14
%
r
h
3 Karpur, Frasanna, and Reach, Michael T. , "Improved Detectability
of Fatigue Microcracks by Split Spectrum Processing of
Backscattered Rayleigh Waves", Review of Progress in Nondestructive
Evaluation. Vol. 10, D.O. Thompson and D.E. Chimenti, eds. , Plenum
Press, New York, 1991, in press.
4 Larsen, J.M., Jira, J.R. , and Weerasooriya, T., "Crack Opening
Displacement Measurements on Small Cracks in Fatigue", Fracture
Mechanics: Eighteenth Symposium, ASTM STP 945, D.T. Read and R.P.
Reed, Eds., American Society for Testing and Materials,
Philadelphia, 1988, pp. 896*912.
5 Karpur P., Shankar, P.M., Rose, J.L. and Newhouse, V.L., "Split
Spectrum Processing: Optimizing the Processing Parameters Using
Minimization", Ultrasonics. Butterworth Scientific, Vol. 25, July
1987, pp. 204-208.
6 Karpur P, , Shankar, J.L. , Rose, J.L. and Newhouse, V.L., "Split
Spectrum Processing: Determination of the Available Bandwidth for
Spectral Splitting", Ultrasonics. Butterworth Scientific, Vol. 26,
July 1988, pp. 204-209.
117-15
1990 USAF-UES SUMMER TACULTY RESEARCH PROGRAM/
GRADUATE STUDENT RESEARCH PROGRAM
Sponsored by the
AIR FORCE OFFICE OF SCIENTIFIC RESEARCH
Conducted by the
Universal Energy Systems, Inc.
urn BfPQRI
NUB sod i£ Investigations si Conformational Dynamlg.?.
iQd ?vr.f3C9 intfiMctlona ol Pgrf 1 tiprwplJYal tori
Prepared by:
Academic Rank:
Department and
University:
Research Location:
Martin Schwartz
Professor
Chemistry Department
University of North Texas
Materials Laboratory
WRDC/MLBT
Wright-Patterson AFB
Ohio 45433-6533
USAF Researcher:
Date:
Contract No:
Dr. Kent J. Eisentraut (Technical Focal Point)
and Dr. Harvey L. Paige
September 13, 1990
F49620-88-C-0053
NMR and IB Investigations of Conformational Dynamics
and Surface Interactions fif Perf 1 uoropoTyal Icvl ethers
by
Martin Schwartz
ABSTRACT
Fluorine-19 NMR spin-lattice (T,) relaxation times were measured
for several perfluoropolyalkyl ethers (PFPAE's). Derived rotational
correlation times (t^) revealed that perfluoromethylene (CFj) groups
adjacent to -OGgF^O- chain segments rotate more slowly than those
attached to -OCFjO- fragments. The decreased chain mobility was
investigated using molecular mechanics to model bond rotations in linear
PFPAE's. The calculations revealed that -OC2F4O- units introduce
steric repulsions which severely restrict rotation about neighboring C-0
bonds. These results can be used to explain the generally observed
correlation between C:0 ratios and fluid viscosities in perfluoroethers.
Preliminary semi -empirical quantum mechanical calculations of
conformational energies and potential barriers have been performed for
several perfliiorocompounds. The results will be compared with those
from ali initio computations and, when available, to experimental data.
The ultimate goal of these studies is tp develop realistic
conformational potential energy functions, which will permit the
prediction of static and dynamic fluid properties and, therefore, aid in
the design of hew PFPAE lubricants..
The application of infrared microscopy to characterize the
interactions of fluid additivas with metal surfaces was investigated.
Several problems in the acquisition of reliable spectral data were
noted. It is recommended that further tests be performed in conjunction
with XPS experiments to provide a definitive assessment of the utility
of IR spectroscopy in the study of chemisorbed additives.
ACKNOWLEDSEMENTS
I would like to thank the Air Force Systems Command and the Air
Force Office of Scientific Research for sponsorship of this research,
and Universal Energy Systems for the able administration of the Summer
Faculty Research Program.
I am indebted to Harvey Paige, with whom I worked on molecular
modelling of conformational energies, and to Jim Liang, my
co-investigator in the IR microscopy investigations. Without their help
and collaboration, I would have made far less progress in research this
summer. I am most grateful to Ed Snyder, Kent Eisentraut and Lois
Gschwender. Their doors were always open to answer my questions and
provide help in any area I requested. They have all become friends and
colleagues and I look forward to continue working with them in the
future.
Finally, I would like to thank all of the members of MLBT, whose
friendship and help »iade my stay at the Materials Laboratory a most
enjoyable and enriching experience.
I. INTRODUCTION
Perfluoropolyalkyl ethers (PFPAE's) are a class of liquids
possessing many physical and chemical properties desirable in a liquid
phase lubricant, including a wide liquid range, excellent thermal
and oxidative stability, a high viscosity index, good lubricity and
shear stability, and, importantly, they are almost completely
non-flammable. No currently available commercial PFPAE lubricants,
however, are capable of operation at the temperature extremes (-54 °C to
+371 “C), in an oxidative environment, required for lubrication of a
high performance gas turbine jet engine. The Air Force has begun a
multi-year interdisciplinary program (IHPTET)^ to study the physical
and chemical properties of PFPAE's in order to develop suitable high
temperature Jet engine lubricants and additives.
I have extensive experience in the application of Nuclear Magnetic
Resonance (NMR) and Fourier Transform Infrared (FTIR) spectroscopy to
study the reorientational dynamics and intermolecular interactions of
molecules in the liquid phase. Both NMR arid FTIR spectroscopy are
potentially quite useful for the investigation of the liquid phase
structure of lubricants and their interactions with surfaces.
Therefore, I was invited to participate in the high temperature
lubricant research program at the Materials Laboratory at
Wright-Patterson AFB this summer.
II. OBJECTIVES OF THE RESEARCH EFFORT
During my pre-summer visit to the Materials Laboratory, Dr. Kent
Eisentraut (my technical focal point), Ed Snyder, and I determined that
I would pursue two research goals this summer: (1) explore the
application of Fluorine-19 NMR spectroscopy to study the liquid phase
three dimensional structure (conformation) and polymer chain dynamics of
PFPAE lubricants and (2) assess the utility of FTIR microscopy to
profile the interact'ions of lubricants and additives with metal
surfaces. The NMR spectra were to be acquired by the graduate students
in my research group at the University of North Texas on samples
provided by the Materials Laboratory. The latter FTIR investigations
were to be performed on site in collaboration with Dr. Jim Liang, ai
research chemist with the University of Dayton Research Institute, who
is permanently assigned to the Materials Laboratory.
During the course of the summer, it became clear that quantum
mechanical and classical molecular modelling, using programs available
at the Materials Laboratory offered an excellent approach, complementary,
to NMR, to investigate the conformation and chain dynamics in liquid
lubricants. Work in this area was begun in the latter half of the
summer in collaboration with Dr. Harvey Paige of the Materials
Lahore tory.
III. NMR RELAXATION AND CONFORMATIONAL DYNAMICS
Introduction The measurement of Nuclear Magnetic Resonance
spin-lattice relaxation times (T/s) is a well established technique to
probe both the rates and mechanism of molecular reorientation in
liquids.^ The method has also been used quite profitably to
characterize the conformational mobility of flexible chain polymers.®'^
Almost all of the studies to date have been on carbon- 13 or proton
relaxation, commonly in hydrocarbon, halocarbon or alkyl ether polymer
chains. During the summer, we investigated the applicability of
fluorine-19 NMR relaxation times to characterize the chain
reorientational dynamics in PFPAE's of varying molecular structure. We
have also begun studies on the utility of a new two dimensional NMR
technique (2D-N0ESY spectroscopy) to furnish information on the
equilibrium conformational behavior of these molecules.
Experiments Fluorine- 19 NMR measurements were performed by
students in my research group at the University of North Texas on a
Varian VXR-300 Fourier Transform NMR Spectrometer operating at
Vo{^’n*282 MHz (Bg=70.5 kG). Spin-lattice (T,) relaxation times were
determined using the standard Inversion Recovery pulse sequence/
(180®^T-90°-Acq.)n> with 10-12 t values plus t-»«. T, was calculated from
the peak intensities by a non-linear fit to the three parameter
magnetization equation.®
Measurements were performed on the following liquids (at ambient
temperature, 25 °C): (1) Perfluoropoly (ethylene oxide), CF30[CF2CF20]„CF,
(ML088-50); (2) Perfluoropoly(triethylene glycol), R^0[(CF2CF20)3CF20]„R^
• CF3, C2F5 (ML088-131); (3) Krytox-AC, C3F70[CF(CF3)CF20]„C2F5
(ML071-6); (4) Fomblin-Z, R^0[CF203JC2F,0]„[C3F,0]qR^, R^-CF3, C2F5
(ML078-80). In addition, two dimensional COSY’'’®'”'’® and
NOESY’®,’®'’^ spectra were acquired for Krytox-AC and Fomblin-Z using
standard techniques.’®*’®
Results The ”f NMR results for the five compounds above are
presented in Tables 1-4 of Appendix A. As expected, the NMR spectra of
Krytox and Fomblin are more complex than those of the model compounds
(ML088-50, ML088-100, ML088-131). A partial set of assignments for
these two lubricants are given in the tables. We have acquired 2D COSY
spectra for these liquids which, together with NMR spectra of a second
molecular weight fraction (to permit assignment of end groups), to be
obtained this fall, will permit us to complete their assignments.
The predominant relaxation mechanism for fluorine nuclei in large
molecules is via dipolar interactions with neighboring ”F spins.
Therefore it is possible, using standard relations,’® to calculate a
quantity termed the correlation time, x^., which, for a CF2 unit in a
perfluoroether, represents (approximately) the time for the vector
between the two fluorine nuclei to rotate by one radian (57°). Thus,
Tp, is a measure of the degree of chain mobility (flexibility) in the
region of the CFj group.
Several trends have emerged from the data acquired this summer.
In all cases, rotational times for CF3O- end groups are very short,
118-6
verifying that there is no significant barrier to rotation of this group
about the C-0 axis. Reorientation of the perf 1 uoroniethyl group in
CFjCFjO- units is somewhat slower {longer t^), as a result of the
three-fold barrier to rotation about the C-C bond. One finds, too, from
the tables that rotational times (t^) of CFj groups within three atoms
of the chain ends are comparatively short, a feature generally observed
in polymer dynamics,^'* since there is a lowered frictional torque
retarding bond rotation in this region of the chain.
The most significant trend in the data analyzed to date may be
seen in the correlation times of Perfluoropoly(tri ethylene glycol) (A. 2)
and Fomblin (A. 5). Without exception, perfluoromethylene (CFj)
rotational correlation times are shorter in -CF^QCF^O- units than in
-S£20C2F40- fragments on the chain. In order to understand this
observation, we utilized a molecular mechanics modeling program (CHEM-X)
to simulate rotation about the two central C-0 bonds in a
-OCFjCFj-O-CFj-O- segment of a perfluoropolyether chain. From the
resulting potential energy contour plot, we observed that rotation of
the right hand -OCFjO- fragment about the first C-0 bond is severely
restricted due to steric repulsions between this CFj group and the
-CFjCFj. unit on the left hand side of the chain segment. In contrast,
the three-fold barrier to rotation of the left hand -OCFjCFjO- fragment
about the second C-0 bond is relatively low, resulting in more rapid
rotation and, thus, shorter correlation times for this group. We note
that the lowered rotational mobility caused by the presence of -OCjF^O-
(and -OCjF^O-) segments in the polymer chain can explain the generally
observed positive correlation between C:0 ratios and viscosities in
PFPAE's.
NMR relaxation times measure the rate of change of a polymer's
conformation via rotation about its various bonds. The equilibrium (or
average) conformation also exerts a major influence on the molecule's
bulk and molecular properties. Recently, two dimensional NOESY
118-7
experiments have been applied to study proton. . .photon distances and,
hence, equilibrium structures in photeins^^'^® and other polymers.’’
We are trying to extend this new experimental technique to obtain
quantitative fluorine*** fluorine distances (Which are a sensitive
function of the average dihedral angles about the C-C and C-0 bonds) in
PFPAE's. During the summer, we have acquired initial fluorine NOESY
spectra in Fpmblin-Z. Further refinement of the spectral acquisition
parameters, necessary to obtain quantitative intensities and F***F
distances, is currently in progress.
IV. MOLECULAR MODELLING OF CHAIN CONFORMATION AND MOBILITY
Introduction In order to understand and, eventually, to predict the
equilibrium conformations and chain mobilities in PFPAE's, it is
essential to have an accurate knowledge of the potential energy as a
function of rotation angle about various single bonds in these fluids.
Once the potential energy functions have been accurately characterized,
it is possible to utilize well-established statistical mechanical
methods^® to calculate both static and dynamic properties as a function
of temperature in the liquid phase.
During the latter half of the summer, I began collaborative
studies with Dr. Harvey Paige of the Materials Laboratory to investigate
the applicability of afe initio^’ and semi -empiri cal quantum
mechanical techniques and of classical molecular mechanics to determine
conformational potential energy functions in PFPAE's.
Calculations and Results As discussed in a previous section, the
classical molecular mechanics model was employed successfully to explain
the trend in rotational correlation times obtained from NMR relaxation
times in Fomblin-Z. During the summer, we also utilized this simple
model to generate energy contour maps for rotation about the two central
C-0 bonds in chain segments of a linear (-OCFj-O-CFjO-) and branched
(-0CF2-0-CF(CF3)0-) perfluoroether. The results revealed that the
118-8
branching perf 1 uoromethyl group in the latter segment induces, steric
repulsions which severely restrict the rotation about the right hand
0-CF(CF3) bond. The resulting reduced chain mobility explains,
qualitative, the general observation that branched PFPAE's exhibit an
order of magnitude greater viscosity than linear perfluoroethers of the
same molecular weight.
We also began preliminary calculations on the conformational
energies of several simple perfluorocompounds (Appendix B), using three
of the semi -empirical quantum mechanics programs (MNDO, AMI and PM3) in
the MOPAC^^ molecular orbital package. The three methods differ
primarily in their integral parametrization by comparison to
experimental results. As a consequence of their comparative scarcity,
there is relatively little data on perfluorocompounds, with which to
parametrize the semi -empirical Hamiltonian. Our goal in this portion of
the project is to determine which of the three Hamiltonians yields
conformational energies in closest agreement with afe initio calculations
and, where available, with experimental conformational energy
differences, potential barriers and molecular dipole moments.
Perfluorodimethyl ether (CF3OCF3) is one molecule whose structural
parameters have been determined experimentally,^^ and for which ab
initio calculations have been reported in the literature. As seen in
Table B.l, the PM3 Hamiltonian yields bond lengths and angles that,
generally, are in superior agreement with experiment than those from
either AMI or MNDO, or, interestingly, from thij ab initio calculation
using a minimal basis set (ST0-3G). The results from PM3 agree
reasonably well, too, with parameters calculated with the extended 4-316
basis functions.
The hypothetical molecule, tetrafluoroethane-l,2-diol
(HOCFjCFgOH) , like perfluoroethers, contains polar C-F and C-0 bonds,
whose interactions influence the relative energies of equilibrium
conformers and potential barriers. We have used the three
118-9
semi -empirical Hamiltonians {PM3, AMI and MNDO) and classical molecular
mechanics (MM2) to calculate the energy arid dihedral angle of the
equilibrium gauche conformer (Eg) and of the two barriers (Egg and Eg^),
relative to E^aO. One sees from Table B.2 that the energies from the
various methods differ substantially, by as much as a factor of 3-4.
Work is in progress to determine which of these methods yields results
in closest agreement with an extended basis set ib initio calculation.
The conformational potential energy function has also been
calculated for perfluorobutane (CFjCFjCFjCFj) using the semi -empirical
methods. As found above, the three Hamiltonians yield markedly
different results'^ (Table B.3). Here, too, ab initio calculations are in
progress. We also intend to obtain perfluorobutane and measure the
relative equilibrium energies of the gauche and trans conformers
experimentally (by temperature dependent IR intensity ratios) as well as
its gas phase dipole moment. A comparison with the predictions of the
various methods will provide important evidence on their relative
merits.
V. IR MICROSCOPY AMO SURFACE INTERACTIONS
Introduction To improve their performance under extreme conditions,
most lubricants require the addition of small amounts of various
additives. One current theory of the action of these additives is that
they chemisorb on an engiie's metal surfaces, passivating the metal from
corrosion, oxidation and wear. FTIR microscopy^® has proven to be a
useful new technique to study the adsorption of molecules on
surfaces.^* Since one can monitor a region as small as 10-20 microns
in diameter, it is potentially possible obtain a detailed profile of
additive/surface Interactions on scarred or corroded regions of metal.
Too, the infrared spectra may be able to provide more detailed
information on the structure of chemisorbed species than is possible
from most other surface analysis techniques. One goal of my summer
118-10
research at the Materials Laboratory was to assess the capability of
FTIR microscopy to study chemisorption of additives on metal surfaces
that had been exposed to four ball wear and oxidation/corrosion tests.
This work was performed in collaboration with Dr. Jim Liang of the
University of Dayton Research Institute.
Experiments IR spectra were obtained on. a Perkin-Elmer 1750
Fourier Transform Infrared Spectrometer, equipped with a Spectra Tech
IR-Plan Infrared Microscope (Model no. 0043-033) and MCT detector. The
following tests were run: (1) Four ball wear test on 5P4E
(5-phenyl -4-ether) lubricant without additive, (2) Four ball wear test
on 5P4E with 5% TCP (tricresyl phosphate) antiwear additive, (3) washed
balls from earlier four ball wear test, (4) Oxidation/Corrosion test on
a branched PFPAE (ML089-293), and (5) Reflection tests of unused
lubricants on metal coupons.
Results Three problems were encountered in the application of IR
microscopy to study additive/metal interactions: (1) similarity of
lubricant and additive IR spectra, (2) spectral anomalies caused by
reflection off uneven metal surfaces, and (3) sensitivity; each is
addressed in turn.
(1) Similarity of lubricant and additive spectra. Standard organic
additives lack sufficient solubility to be useful in improving the
performance of the new high temperature lubricants. Therefore, it has
been necessary to synthesize new additives with structures very similar
to that of the lubricant in order to enhance solubility. As a
consequence, one usually finds that the IR of the additives and base
fluids overlap in all spectral regions. Thus, it is quite difficult to
monitor changes in the additive's spectrum if any lubricant remains on
the metal .
(2) Spectral anomalies. It is well known that IR reflection spectra of
films on metal surfaces depend critically upon the film thickness and
reflection angle;^^'^® neither thickness nor angle can be controlled
118-11
on scarred or corroded metal surfaces. As a result, we observed that
the relative peak intensities and positions of adsorbed film spectra
varied irregularly from one experiment to the next, even when unused
(and thus unaltered) lubricants were applied to clean balls or metal
coupons. Therefore, one must be extremely cautious in ascribing any
variations in the IR spectrum to chemical modification resulting from
surface adsorption.
(3) Sensitivity. Ideally, one may eliminate the above problems by
washing the metal surface to remove any residual lubricant film on the
surface, leaving only the chemisorbed additive. This procedure has been
reportbu to be successful in obtaining the XPS spectrum of adsorbed
species. A number of experiments were performed on washed balls and OC
coupons. Unfortunately, in none of the tests were we able to observe
any signal due to chemisorbed species.
VI. RECOHNENOATIONS
NMR Relaxation As outlined above, the results obtained to date
demonstrate that Fluorine- 19 NMR T/s can provide valuable information
on the effect of the local molecular environment on the chain mobility
in PFPAE's. Investigations of additional model compounds and of the
effect of medium (temperature, solvent, polymer mixtures...) are
currently underway. In addition, we intend to begin measuring carbon-13
T/s and spin-spin (Tj) and rotating frame (T^^,) relaxation times to
obtain complementary information on the chain dynamics. I am currently
preparing a research proposal for an RIP grant to request funding for
this continuing research.
Molecular Modelling Although calculated parameters obtained from the
PM3 semi -empirical Hamiltonian were in good agreement with the
experimental values, there is insufficient data to ascertain whether
this method is generally superior to AMI and MNDO. As further ab initio
and experimental results are accumulated, it may prove appropriate to
118-12
develop a new semi -empirical Hamiltonian designed specifically for the
perfluoroethers.^’
The simple molecular mechanics model has proven useful to
ascertain qualitative trends in conformational energies of
perfluoroethers (vide suoral. However, the results bear no quantitative
significance. As reliable quantum mechanical and experimental results
become available, it will be possible to introduce realistic torsional
potentials and bond stretching and bending force constants into a
modified molecular mechanics model parametrized specifically for
PFPAE's. It will then become feasible to utilize this very fast
computational procedure to perform molecular dynamics calculations on
the PFPAE's. The results can be used to predict rotational correlation
times, which can be compared to those measured experimentally from NMR
relaxation to determine the adequacy of the modified classical model.
Once accurate potential functions have been developed, they may be used
to calculate both static and dynamic properties of these fluids, thus
aiding in the design of new PFPAE lubricants.
IR Microscopy It is apparent from the problems described in
Section V that reliable IR spectra of chemisorbed additives cannot be
determined in the presence of residual lubricant fluids due to their
spectral similarities and the anomalous intensities and peak positions
of reflectance spectra from liquid films. As noted, when the balls
(from four ball experiments) or OC coupons were washed, no IR spectra of
adsorbed species were detected. However, rather than resulting from a
lack of sufficient sensitivity, it is possible these metals contained no
significant amounts of the additive. We (Dr. Liang and I) recommend that
when XPS studies on these systems are begun in the coming year, that IR
spectra be recorded at the same time. A comparison of results from the
two experiments should provide a definitive test of the utility of
infrared microscopy to characterize the interactions of additives with
metal surfaces.
118-13
REFERENCES
1.. C. E. Snyder, Jr. and R. E. Dolle, Jr., ASIE Trans. 19, 171 (1975)
.2. C. E. Snyder, Jr., L. J. Gschwender and C. Tamborski, Lubr. Eng. 37.
344 (1981)
3. Integrated High Performance Turbine Engine Technology (IHPTET)
Program
4. R. T. Boere and R. G. Kidd. Ann. Reo. NMR Soectrosc. 13, 3i9 (1982)
5. F. Heatley, Prog, in NMR Soectrosc. 13, 47 (1979)
6. F. Heatley, Ann. Reo. NMR Soectrosc. 17, 179 (1986)
7. M. L. Martin, G. J. Martin and J.-J. Delpuech, Practical NMR
SbectroscoDv. Heyden, London (1980), Chap. 6
8. A. A. Rodriguez, S, J. H. Chen and M. Schwartz, J. Magn. Reson. 74,
114 (1987)
9. COSY is the acronym for Correlation Spectroscopy
10. W. P. Aue, E. Bartholdi and R. R. Ernst, J. Chem. Phvs,. 64, 2229
(1976)
11. A. Bax and R. Freeman, J. Maon. Reson. 42, 181 (1981)
12. R. Benn and H. Giinther, Anoew. Chem. Int. Ed. Engl. 22, 350 (1983)
13. NOESY is the acronym for Nuclear Overhauser Enhancement Spectroscopy
14. B. iH. Meier and R. R. Ernst, J. Am. Chem. Soc. 101, 6441 (1979)
15. G. Wider, S. Macura, A. Kumar, R. R. Ernst and K, Wuthrich, J. Maon.
Reson. 56, 207 (1984)
16. E. D. Becker, High Resolution NMR: Theory and Applications. 2nd;.
Ed., Academic, New York (1980), Chap. 9
17. A. Kumar, G. Wagner, R. R. Ernst arid K. Wuthrich, J. Am. Chem. Soc.
103, 3654 (1981)
18. M. Weiss^ D. Patel, R. Sauer and M. Karplus, Proc. Natl. Acad. Sci.
USA 81, 130 (1984)
19. P. A. Mirau, F. A. Bovey, A, E. Tonelli and S. A. Heffner,
Macromolecules 20, 1701 (1987)
20. P. J. Flory, Statistical Mechanics- of Chain Molecules.
Wiley- Interscience, New York (1969)
118-14
21. W. J. Hehre, L. Radom, P. von R. Sch.leyer and J. A. Pople, Ab Initio
Molecular -Orbital Theory^ Wiley, New York (1986).
22. J. J. P. Stewart, J. Computer-Aided Molec. Design 4, 1 (1990)
23. J. Pacansky and B. Liu, J. Phvs.Chem. 89, 1883 (1985)
24. A. H. Lowrey, C. George, P. D'Antonio and J. Karle, J. Mol. Struct.
63, 243 (1.980)
25. R. G. Messefschmidt and M. A. Harthcock, Infrared Microspectrometrv;
Theory and Applications. Marcel Dekker, New York (1988)
26. J. F. Rabolt,. M. Jurich and J. D. Swalen, Add!. Soectrosc. 39, 269
(1985)
27. D. L. Allara, A. Baca and C. A. Pryde, Macromolecules 11, 1215
(1978)
28. V. J. Novotny, I. Hussia, J.-M. Turlet and M. R. Philpott, J. Chem.
Phvs. 90, 5861 (1989)
29. Dr. James J. P. Stewart, Project Scientist at the Frank J. Seiler
Research Laboratory, was a principal participant in the development
of the MOPAC molecular orbital package, and has, himself, developed
the newer PM3 extension. Dr. Stewart, in correspondence with Dr.
Harvey Paige, has indicated his willingness to develop a new
semi -empirical parameter set specifically for perfluorocompPunds, as
ab initio and experimental results become available.
APPENDIX A. NNR RELAXATION AND CORRELATION TINES
A.l pepfluoropoly(ethylene oxide) , CF30tCF2CE20j„CF3 (MLOSS^SO)
Peak
Delta*
Assignment
h
1
73v6 ppm
CFjOCFjCFjO
0.80 s
150 ps
2
75.8
OCF2CFP
0.47
250
3
108.3
CF3OCF2CF2O
1.09
60
♦Chemical shifts in this and following tables are measured in ppm
doWnfield from hexafluorobenzene.
A.2
Perf 1 uoropol y ( tri ethyl ene gl ycol )
Rf « CF3, CjFj (HL088-131)
, R,OI(CF2CF20)3CF20]„R^ ,
Peak
Delta
Assignment
Tr
1
73.7 ppm
CF3OCF2CF2O
0.81 s
140 ps
2
74.0
OCF2CF2OCF2O
0.63=
180
3
75.7
0C2F40CF2CF20C2F^0
0.59
200
4
76.8
CF3CF2O
0.60
110
5
108.2
CF3OC2F4O.
1.21
60
6
112.6
PC2F40CF20C2F40
0.67
210
118-16
A.3 krytox-AC , C3F7Q[CF(CF3)CF203„C2F5 (ML071-6)
Peak
Delta
Assignment
1
19.7
0CF2CF(CF3)0
0;26
130 ps
2
20.0
0CF2CF{CF3)0
0.20
130
3
20.3
OCF2CR{CF3)0
0.20
130
4
34; 4
CFjCFjCFjO
0.36
280
5
75.6
OCF2CF3
0.50
130
6
76.8
CF3CF2CF2P
0.49
240
7
82.4
CFsCFzCFjO
0.52
120
8
83.8
0CF2CF{CF3)0*
0.40
290
9
84.2
0CF2CF(CF3)0
0.38
170
10
84.3
0CF2CF(CF3)0*
0.38
310
11
84.8
0CF2CF{CF3)0*
0.40
290
12
85.3
0CF2CF(CF3)0*
0.37
320
*The CFj fluorines in Krytox are non-
to an asymmetric carbon. Therefore,
large geminal coupling constant
equivalent since they are adjacent
they split into a quartet with a
t
U8-17
A, 4
FbmbTin-Z ,
Rf^CFji
C2F5 (NL078-80)
Peak
Delta
Assignment
Ti
1
35.2 ppn
1 OCFjCFgCFjO
0.52 s
200 ps
2
74.2
pCFjCFjOCFzQ
0.63
190
3
75.9
OCFjCFjOCjF^O
0.51
230
4
79.5
dCFjCFjCFjOCFjO
0.63
190
5
81.1
0CF2CF2CF20C2F«0
0.50
230
6
106.9
CF3OCF2O
1.08
70
7
108.6
CFjOCjF^O
1.19
60
8
109.5
OCF2OCF2OCF2O
0.95
140
9
111.3
0CF20CF20C2F^0
0.78
180
10
111.5
CFjOCFjO
0J9
170
11
112.9
OC2F4OCF2OC2F4O
0.66
210
118-18
APPENDIX B. CALCULATED CONFORNATIONAL ENERGIES
B.I Perfluorodi methyl ether (CF3OCF3): Comparison between theory
and experiment
Param
Expt."
4-316^*
ST0-3G‘’
PM3'
AMI"
MNDO"
C-O*'
1.369
1.369
1.431
1.392
1.406
1.406
C-F*^
1.327
1.340«
1.366“
1.334“
1.353“
1.344“
C-O-C*
119.1
125.9
113.2
118.9
121.3
130.4
F-C-O-C^
166.0
162.2
162.0
161.5
165.8
165.2
a) Ref. 23
b) Ref. 24
c) This work
d) Bond length (in Angstroms)
e) Bond angle (in degrees)
f) Dihedral angle (in degrees)
g) Average value
B.2 Tetrafluoroethane-l,2-diol (HOCFjCFjOH)*
Method
E**
E®
•-GC
*GG
••GT
*GT
PM3
0.24
62
2.98
0
2.25
119
AMI
0.59
80
1.83
0
0.69
112
MNOO
0.53
64
4.15
0
3.21
120
MM2
0.27
62
6.41
2
5.13
120
a) All energies given in kcal/mol, referenced to Ey=0
b) Energy and dihedral angle of gauche (G+ and G*) equilibrium
conformations
c) Energy and dihedral angle of barrier between G+ and G' conformations
d) Energy and dihedral angle of barrier between G+ (or G’) and T
conformations
118-19
B.3 Semi -Empirical Conformational Energies of Perfluprobutane*
PM3
AMI
MNDO
0“
5.89 kcal
4.60 kcal
8.01 kcal
10
6.47
4.52
7.54
20
6.74
6.41
30
2.72
3.211
4.99
40
1.17
2.62
3.60
50
-0.36
2.00
2.57
60
-0.08
1.53
2.11
70
1.29
1.28
2.19
80
1.62
1.16
2.37
90
0.93
0.81
2.13
100
0.25
0.62
2.06
110
0.98
0.59
2.23
120
1.43
0.56
2.27
130
1.42
0.47
2.00
140
1.24
0.31
1.45
150
0.47
0;14
0.80
160
-0.23
0.01
0.28
170
0.07
-0.02
0.04
180
0.00
0.00
0.00
a) All energies in kcal/mol, and referenced to E(180“)=0
b) Dihedral angle between C, and
118-20
1990 USAF-UES SUMMER FACULTY RESEARCH PROGRAM/
GRADUATE STUDENT RESEARCH PROGRAM
Sponsored by the
AIR FORCE OFFICE OF SCIENTIFIC RESEARCH
Conducted by the
Universal Energy Systems, Inc.
FINAL REPORT
MODELING OF CASTING SOLIDIFICATION
Prepared by:
Academic Rank:
Department and
University:
Research Location:
USAF Researcher:
Date:
Contract No:
Hai-Lung Tsai
Assistant Professor
Mechi and Aero. Engr. and Engr. Mech.
University of Missouri^Rolla
WRDC/MLLM
Wright-Patterson AFB, OH 45433
James C. Malas
September 5, 1990
F49620-88-C-0053
A general purpose finite element computer program, CAST3,
for modeling casting solidification was evaluated from both the
user's and the technical aspects. The CAST3 code was developed
by the Universal Energy Systems, Inc. under the sponsorship of
the Air Force. Although several commercial packages are avail¬
able, it was found that CAST3 is the only software dedicated to
the casting solidification modeling. As a result, CAST3 code
has a superior capability in handling the casting-mold interfa¬
cial thermal resistance and the time stepping algorithm, which
make the program computationally more efficient than any other
available codes (to the knowledge of the author) . An excellent
start has been made by CAST3 in achieving the goal of develop¬
ing an ideal casting design package for the Air Force. How¬
ever, the present version of CAST3 is not yet completed for
being able to simulate some casting problems. Therefore, rec¬
ommendations are made for improving and expanding CAST3, so
that the code can be used in the foundry industry as a powerful
design tool for obtaining high quality casting parts.
I. INTRODUCTION
Metal casting foundry is a basic industry which provides
either raw materials or finished parts for many other manufac¬
turing industries. Hence, the casting production technology in
the foundry will affect directly or indirectly the cost as well
as the quality of many downstream products. During the past
years, the production of American castings has been decreased
significantly, due to the loss of market to foreign countries
where the labor is much cheaper [1]. One of the major issues
for the American foundry to survive and maintain the leading
edge in the global competition is the introduction of modern
technology to this industry for obtaining castings of better
quality at a lower cost. To achieve this, the use of computer
modeling as well as CAD/CAM in the foundry are required.
To date, several commercial finite element packages are
available for modeling casting solidification, which include
ABAQUS, ADINA, ANSYS, NASTRAN, NISA, MARK, etc. In general,
these programs were developed for the general purpose use in
many areas, and they are not dedicated to the metal casting
simulation. As a result, special features unique to casting
solidification, such as the latent heat release and the metal-
mold interfacial thermal resistance are not well-handled in
these programs, which crucially affect the required computer
time. Consequently, for any of the above-mentioned code, it
needs a tremendous computer time to simulate a real casting
process, which hinders the usage of computer modeling in the
119-3
foundry industry.
During the past years, Drs. Mark Samonds and David Waite of
the Universal Energy Systems, Inc. have developed a general
purpose finite element code, CASTS, dedicated to the modeling
of casting solidification, under the sponsorship of the Air
Force. It has been demonstrated by several foundries that the
basic portion (heat conduction) of CASTS is very useful in pre¬
dicting the transient temperature distributions in the casting
and the mold, so that any possible hot spots in the casting
could be avoided by changing some casting conditions. However,
so far a rigorous evaluation on CASTS, from either a user's
point of view or the technical aspect, is not available. The
author has worked on the casting solidification modeling for
about 10 years, including the use of several above-mentioned
commercial codes and many years of actual foundry experience.
In addition, he has developed part of NISA code, and is famil¬
iar with both finite element and finite difference methods.
Hence, the author is one of the very few candidates who are
qualified for evaluating CASTS code.
1194
My assignment as a participant in the 1990 Summer Faculty
Research Program was to evaluate the code CAST3, and then to
make recommendations on the direction for further research in
the modeling of casting solidification. The Air Force was par¬
ticularly interested in the D'arcy flow portion in the CAST3.
Hence, it was determined to run CAST3 for a unidirectional
macrosegregation problem, so that CAST3 can be validated by
comparing the computational results with the available exper¬
imental data.
The CAST3 User Manual was studied in detail and a one
dimensional solid model was established using PATRAN for simu¬
lating the macrosegregation problem. As many places in the
manual were not clearly described, the CAST3 program developer.
Dr. Mark Samonds, was called to discuss the technical contents
of CAST3. In order to be able to exchange ideas more conve¬
nient, Mark was invited to the laboratory. The discussion has
gone through every phase of the program CAST3. It was found
that the D'arcy flow portion in CAST3 was developed under the
assumption that a continuous feeding of molten metal is pro¬
vided, which is inconsistent with an actual macrosegregation
experiment, in which there is no additional feeding. It was
also found that the proposed macrosegregation problem has been
simulated using CAST3 by Dr. David Waite, and the results were
available. Although the modeling results appeared reasonable
based on the assumption made when the program was developed,
they were completely different from the experimental data. It
was concluded that a major modification in CASTS is required in
order to be able to predict the macrosegregation phenomenon
observed in experiments. As a result, the idea of using CASTS
to simulate the unidirectional macrosegregation problem was
abandoned, and efforts were focused on identifying the merits
and limitations of CASTS, as well as the directions for pos¬
sible improvements on CASTS.
III. MERITS AND LIMITATIONS OF CAST3
The most distinctive difference between CAST3 and other
commercial programs is that CAST3 was developed solely for
modeling casting solidification. Hence, special features
unique to casting solidification are well-handled in CAST3.
The merits of CAST3 are summarized as follows:
MERITS:
1. Method in handling metal-mold interfacial problem
The technique used in CAST3 to handle the casting-mold
interfacial thermal resistance is excellent. In casting model¬
ing, the domain consists of both the casting and the mold,
which is a conjugate problem. It is noted that usually the
thermophysical properties of the casting are very different
from those of the mold. For example, the thermal diffusivity
of aluminum based alloys is about 100 times that of a silica
sand mold. Especially, there is a phase change in the casting
which introduces nonlinearity in the numerical solution. The
technique used by CAST3 can have a complete different grid sys¬
tem and time stepping scheme for each of the casting and mold
domains, so that the computational efficiency can be increased.
The author uses a similar technique to handle the metal-mold
interface, and it is called the "two-domain method” [2].
2. Method in handling latent heat release
In CAST3, the enthalpy method is employed for handling the
latent heat release during alloy solidification. Enthalpy
method has been the most popular technique to deal with latent
heat, even if a fluid flow presents in the casting. The rela¬
tionship between the casting solid fraction and temperature can
be assumed [3], or directly input from the alloy phase diagram.
This provides a more realistic way for implementing the latent
heat in the program.
3* Timg-.gtgpp.ingL,.algari'tebw
In the time domain, depending upon the characteristics of
the problem, CAST3 uses two- or three-level predictor-corrector
time stepping scheme, which gives the user a flexibility. The
scheme can be fully implicit, fully explicit, or in between of
them. Proper selection of the time stepping scheme can sub¬
stantially save the computational time. For example, due to
the large difference in material properties between the casting
and mold, a fully implicit scheme can be used for the casting
domain to guarantee the numerical convergence, while a fully
explicit scheme can be employed in the mold to reduce the com¬
puter time.
4. Computational efficiency
By using CAST3, significant computational cost can be
reduced due to the techniques employed in handling the metal-
mold interfacial thermal resistance and the time stepping algo¬
rithm, as mentioned above. In addition, CAST3 uses the Conju¬
gate Gradient method to solve the assembled matrix, which could
be more efficient than the traditional Gaussian elimination
technique. This is particularly important for a large, sparce
equation systems, which usually occur in a finite element ana¬
lysis. It is expected that CAST3 code can save more than 50%
of computational time, as compared with commercial codes such
as ABAQUS and ANSYS, with which the author is familiar.
5. ga&abilitv in calculating radiation view factor
CAST3 has the capability to calculate the changes of view
factor as a function of time, so that the radiation heat trans¬
fer can be properly accounted. This special feature is partic¬
ularly useful for the simulation of single crystal growth dur¬
ing the pulling process.
LIMITATIONS:
1. Heat conduction only
The major limitation of CASTS is that only heat conduction
in the casting can be handled, but not the fluid flow problem.
The D'arcy flow portion in CASTS was developed under over¬
simplified assumptions, and it is generally not applicable to a
real casting process. As a result, several casting defects
associated with the fluid flow in the casting cannot be pre¬
dicted by CAST3. These include the formations of microsegrega¬
tion, macrosegregation, and porosity in the casting. The fluid
flow in the casting can be caused by thermal and/or solutal
gradients, as well as shrinkage [4].
2. Mold is instantaneously filled
Although theoretically CAST3 allows arbitrary initial con¬
ditions for the casting and the mold, uniform initial tempera-
119-9
tures are usually assumed due to the unknown initial cpndi-?
tions. In order to obtain the initial temperature ahd/or velo¬
city distributions, the mold filling process must be simulated.
In addition, the study of mold filling is important for many
casting processes, for example, the thin-sectipn casting, die
casting, and lost foam casting.
119-10
IV. RECOMMENDATIONS FOR IMPROVING- CAST3
From the above discussion, it can be concluded that CASTS
is an excellent program for modeling casting solidification, if
only heat conduction in the casting is considered. The program
can be employed to determine the constant temperature curves,
so that any possible hot spots in the casting can be detected.
Then, by changing the casting condition a favorable solidifica¬
tion pattern can be achieved and, consequently, hot spots and
the associated casting defects can be eliminated.
However, several improvements in CASTS are required in
order to make CASTS a true tool for predicting and solving sev¬
eral casting problems. A brief discussion of some areas for
possibly including in the future version of CASTS will be given
in the following. However, it is noted that several topics to
be discussed are still under intensive investigation and, per¬
haps, are not matured enough to be implemented in a casting
program such as CASTS.
1. More user-friendlv
The CASTS User Manual needs to add several well-defined and
worked-out examples to illustrate the procedure, step by step,
how to use the program. Tho theoretical development and the
user's manual could be separated into two volumes. There are
many misprints in the manual. It is recommended that the ter¬
minologies used in the manual should be consistent with the
tradition. For example, when there is a fluid flow in thie
casting, the heat transfer occurred is called convection, but
not conduction as used in the manual. Also, several data input
formats are not convenient as comparing with the typical forms
available in the literature. In conclusion, additional efforts
need to be made to enhance the user-friendly of CAST3.
2. Mold filling process
The simulation of mold filling process is important not
only it provides the initial conditions for the subsequent
casting solidification modeling, but also many casting defects
are determined by the filling process. For example, in die
casting, lest foam casting, and injecting molding, the quality
of casting parts is nearly determined by the mold filling pro¬
cess. The most popular technique to handle the free surface
during filling process is described in a program, SOLA-VOF,
developed by the Los Alamos Scientific Laboratory. However,
the original program can handle only the fluid flow portion,
but not the heat transfer. Additional capability needs to be
developed if solidification occurs during the filling process.
3. Fluid flow problem
It is well known that the fluid flow in the casting, due to
either the forced convection, natural convection, shrinkage, or
their combination, can significantly affect the casting
quality. Hence, a casting solidification model must be able to
handle the fluid flow in the casting. The difficulty in hand¬
ling the fluid flow in the casting stems from the fact that a
solidifying casting consists of the solid, liquid, and mushy
119-12
regions. In addition to the vinknown shapes of the interface
between different phases, there is a latent heat release. The
state-of-the-art technique uses a single set of differential
equations throughout the entire casting domain [4].
4. Method for handling mushv region
In the mushy region, both the solid and liquid phases coex¬
ist, hence a special technique need to be developed to handle
the fluid flow and heat transfer in the mushy region. It is
noted that the fluid flow in the mushy region for a solidifying
casting is quite different from the traditional two-phase flow
typically found in the chemical or nuclear engineering. The
state-of-the-art technique is to use the continuum model for
the mushy region, which is based on the volume averaging scheme
[4].
5. Gas- and shrinkage-caused porosities
Casting porosity has been one of the oldest and most con¬
cerned problems in the foundry industry, in particular, for the
aluminium based alloys. The existence of micro pores signifi¬
cant decreases the strength of the casting parts, especially
under the environment at high temperature and high pressure.
The pores in the casting can be caused by either the dissolved
gases in molten metal, the shrinkage, or their combination.
Intensive research on the formation of porosity in casting is
still undertaken both in the academia and the industry [5].
119-13
6. Macroseareaation
For large casting parts,, a severe large-scale macrosegrega¬
tion has been frequently found. It is noted that once the
macrosegregation is created in the casting, there is no means
to eliminate it, except just discard the casting. Macrosegre¬
gation in the casting is caused by the fluid flow both in the
liquid phase and the mushy region. Although some over¬
simplified models on macrosegregation were published in the
early 1960's, a comprehensive mathematical model is still not
available to date.
7. stiEin]aggrin<ay<?g(i(. ghanqg angL.£l\iid..,£,lgw
Although the fluid flow caused by shrinkage is expected to
be very small, it has been found that some casting defects such
as segregation and porosity are mainly owing to the shrinkage
induced fluid flow [6]. This is understandable, since when the
solid fraction in the casting exceeds 50%, the fluid flow in
the mushy zone due to natural convection is almost diminished.
However, the fluid flow caused by shrinkage is always present,
due to the density difference between the solid and the liquid
phases.
8. Actual latent heat release
The use of phase diagram to account for the latent heat
release is possible if there are only two major constitutes in
the alloy. Also, phase diagram is obtained under the assump¬
tion that an equilibrium exists. In any casting solidification
process, some degree of undercooling is always existent in the
119-14
solidifying metal. Hence, the rate of actual latent heat
release should be determined by experiments, so that it can be
directly implemented in the modeling [7].
It is noted that the above-mentioned topics by no means is
a complete list for CAST3 to improve. For example, the stress
built up in the casting during solidification is very important
for determining the hot tear and casting distortion, which, in
turn, can affect the gap formation between the casting and the
The computer program, CAST3, developed by the Universal
Energy Systems, Inc., was evaluated during the Summer Faculty
Research Program. CAST3 code is a unique computer program
dedicated to modeling casting solidification. The code has
many superior capabilities than any existing commercial codes.
However, additional improvements and expansion of CAST3 are
required. It is felt that CAST3 has made an excellent start to
become a sophisticated code, which can be truly used in the
foundry for designing casting processes.
In view of the increasing demands for high quality casting
parts, there is a strong need for a mathematical model, which
can predict and, subsequently, eliminate possible casting
defects. Hence, it is suggested that the Air Force should con¬
tinue to support the research and development in the area of
casting solidification modeling.
119-16
The author wishes to thank the Air Force System Command and
the Air Force Office of Scientific Research for sponsorship of
this research. Thank is also extended to Universal Energy
Systems for their concern and help to me in all administrative
and directional aspects of this program.
The concern and support of Jim Malas was greatly appreci¬
ated. The stimulating discussion with Bill O'Hara, Carl Lom¬
bard, Venkat Seetharaman, and Vinod Jain were really helpful.
The discussion with Mark Samonds has made the present research
possible, which is greatly acknowledged.
REFERENCES
The Census of World Casting Production. Published Yearly
Since 1966 by the American Foundrymen's Society, Inc., bes
Plains, XL.
Chen, J. H. and Tsai, H. L. , "An Efficient and Accurate
Numerical Algorithm for Multi-Dimensional Modeling of Cast¬
ing Solidification - Part I. Control Volume Method," AFS
IranaastiQng/ voi. 98, 1990.
Chen, J. H. and Tsai, H. L. , "Comparison on Different Modes
of Latent Heat Release for Modeling Casting Solidifica¬
tion," AFS Transactions. Vol. 98, 1990.
Chiang, K. C. and Tsai, H. L., "Shrinkage-Induced Fluid
Flow and Domain Change in Two-Dimensional Alloy Solidifica¬
tion," to appear.
Chiou, I. J. and Tsai, H. L., "Modeling on Porosity Forma¬
tion in Castings," AFS Transactions. Vol. 98, 1990.
Chiang, K. C. and Tsai, H. L., "Interaction Between Shrin¬
kage-Induced Fluid Flow and Natural Convection During Alloy
Solidification," to appear.
Su, C. H. and Tsai, H. L. , "A direct Method to Include
Latent Heat Effect for Modeling Casting Solidification," to
appear.
Live Music
Archive
Librivox
Free Audio
Metropolitan Museum
Cleveland
Museum of Art
Internet
Arcade
Console Living Room
Open Library
American
Libraries
TV News
Understanding
9/11